See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/260125327

An Artificial Intelligence Based Finite Element Method

Article · January 2009

CITATIONS READS

6 4,525

4 authors, including:

Akbar Javadi Asaad Faramarzi

University of Exeter University of Birmingham
166 PUBLICATIONS   1,509 CITATIONS    56 PUBLICATIONS   381 CITATIONS   

SEE PROFILE SEE PROFILE

Alireza Ahangar-Asr
University of Salford
28 PUBLICATIONS   220 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Modelling suction caison installation View project

AGRI (Agri-tech Growth and Resources for Innovation) View project

All content following this page was uploaded by Asaad Faramarzi on 11 February 2014.

The user has requested enhancement of the downloaded file.

 ISAST Transactions on Computers and Intelligent Systems, No. 2, Vol. 1, 2009
A. Javadi et.al: An Artificial Intelligence Based Finite Element Method
1

An Artificial Intelligence Based

Finite Element Method
Akbar A. Javadi, Moura Mehravar, Asaad Faramarzi and Alireza Ahangar-Asr

Abstract—In this paper, a new approach is presented based I. INTRODUCTION

S
on artificial intelligence and evolutionary computing, for
imulation techniques, and in particular the finite element
constitutive modeling of materials in finite element analysis,
with potential applications in different engineering disciplines. method, have been used successfully to predict the
This new approach presents a unified framework for response of systems across a whole range of industries
constitutive modeling of complex materials in finite element including aerospace and automotive, biomedical, chemical
analysis using evolutionary polynomial regression (EPR). EPR processes, geotechnical engineering and many others. In this
is a data-driven method based on evolutionary computing, numerical analysis the behavior of the actual material is
aimed to search for polynomial structures representing a
approximated with that of an idealized material that deforms
system. A procedure is presented for construction of EPR-
based constitutive model (EPRCM) and its integration in finite in accordance with some constitutive relationships.
element procedure. The main advantage of EPRCM over Therefore, the choice of an appropriate constitutive model
conventional and neural network-based constitutive models is that adequately describes the behavior of the material plays
that it provides the optimum structure for the material an important role in the accuracy and reliability of the
constitutive model representation as well as its parameters, numerical predictions. During the past few decades several
directly from raw experimental (or field) data. It can learn
nonlinear and complex material behavior without any prior
constitutive models have been developed for various
assumption on the constitutive relationship. The proposed materials. Most of these models involve determination of
approach provides a transparent relationship for the material parameters, many of which have a little physical
constitutive material model that can readily be incorporated in meaning [1]. Despite considerable complexities of
a finite element model. A procedure is presented for efficient constitutive theories, due to the erratic and complex nature
training of EPR, computing the stiffness matrix using the of some materials such as soils, rocks, composites, etc.,
trained EPR model and incorporation of the EPRCM in a
commercial finite element code, ABAQUS. The application of
none of the existing constitutive models can completely
the developed EPR-based finite element method is illustrated describe the real behavior of these materials under various
through two examples and advantages of the proposed method stress paths and loading conditions.
over conventional and neural network-based FE methods are In conventional constitutive modeling, an appropriate
highlighted. mathematical model is initially selected and the parameters
of the model (material parameters) are identified from
Index Terms— Constitutive Modeling, Data Mining, appropriate physical tests on representative samples to
Evolutionary Computation, Finite Elements capture the material behavior. When these constitutive
models are used in finite element analysis, the accuracy with
which the selected material model represents the various
aspects of the actual material behavior affects the accuracy
of the finite element prediction.
In the past few decades, attempts have been made by a
number of researchers to use artificial neural networks
(ANN) to model the constitutive material behavior. The
Manuscript received November 2, 2009.
Akbar A. Javadi is a Senior Lecturer in Geotechnical Engineering in application of ANN for constitutive modeling of concrete
University of Exeter, School of Engineering, Mathematics and Physical was first proposed by Ghaboussi et al. [2]. Ghaboussi and
Sciences, Exeter, EX4 4QF, UK (corresponding author; phone: +44 1392 Sidarta [3] presented an improved technique of ANN
263640; fax: +44 1392 217965; e-mail: [email protected]).
Moura Mehravar is an MSc student in Civil Engineering, Department
approximation for learning the mechanical behavior of
of Civil Engineering in Azad University, South Tehran Branch, Tehran, drained and undrained sand. The role of ANN in
Iran, (e-mail: [email protected]). constitutive modeling was also studied by a number of other
Asaad Faramarzi is a PhD student in Geotechnical Engineering in researchers (e.g., [4-9]). These studies indicated that neural
University of Exeter, School of Engineering, Mathematics and Physical
Sciences, Exeter, EX4 4QF, UK (e-mail: [email protected]). network-based constitutive models can capture nonlinear
Alireza Ahangar-Asr is a PhD student in Geotechnical Engineering in material behavior with high accuracy. While it has been
University of Exeter, School of Engineering, Mathematics and Physical shown that ANNs offer great advantages in constitutive
Sciences, Exeter, EX4 4QF, UK (e-mail: [email protected]).
 ISAST Transactions on Computers and Intelligent Systems, No. 2, Vol. 1, 2009
A. Javadi et.al: An Artificial Intelligence Based Finite Element Method
2

modeling of materials, they also have some drawbacks. One method is illustrated through two examples.
of the main disadvantages of the NNCM is that the optimum
structure of the ANN (e.g., number of layers, number of II. EVOLUTIONARY POLYNOMIAL REGRESSION
neurons in the hidden layers, transfer functions, etc.) must Evolutionary polynomial regression (EPR) is a data-
be identified a priori which is usually obtained using a time driven method based on evolutionary computing, aimed to
consuming trial and error procedure. Another drawback of search for polynomial structures representing a system. A
the NNCM approach is the large complexity of the network general EPR expression can be presented as [18]:
structure as it represents the knowledge in terms of a weight
matrix and biases that are not accessible to user [10]. n

Although one of the main applications of material y = ∑ F ( X , f ( X ), a j ) + a0 (1)

j =1
modeling is in numerical analysis of boundary value
problems, to date not many researchers have considered the
where y is the estimated vector of output of the process;
integration of the neural network based constitutive models
(NNCMs) in numerical modeling techniques such as finite a j is a constant; F is a function constructed by the process;
element method [1]. The main reason for this appears to be X is the matrix of input variables; f is a function defined by
the fact that there are considerable difficulties in the user and n is the number of terms of the target
incorporating a general NNCM in finite element codes [11]. expression. The general functional structure represented by
However, more recently it has been shown that NNCM can F ( X , f ( x ), a j ) + a0 is constructed from elementary
be practically incorporated in a finite element code as a
functions by EPR using a (GA) strategy. The GA is
material model (e.g., [12, 13]). Hashash et al. described
employed to select the useful input vectors from X to be
some of the issues related to the numerical implementation
combined. The building blocks (elements) of the structure
of a NNCM in finite element analysis and derived a closed-
of F are defined by the user based on understanding of the
form solution for material stiffness matrix for the neural
physical process. While the selection of feasible structures
network-based constitutive model [13].
to be combined is done through an evolutionary process, the
Javadi and his coworkers have carried out extensive
research into application of neural networks in constitutive parameters a j are estimated by the least square method.
modeling of complex materials. They have developed an EPR is a technique for data-driven modeling. In this
intelligent finite element method (NeuroFE code) based on technique, the combination of the genetic algorithm to find
the incorporation of a back propagation neural network feasible structures and the least square method to find the
(BPNN) in finite element analysis (e.g., [14–17]). In their appropriate constants for those structures implies some
work, they used actual material test results to extract stress– advantages. In particular, the GA allows a global
strain relationship and to train the NNCM. It has been exploration of the error surface relevant to specifically
shown that NNCMs trained in this way can be very efficient defined objective functions. By using such objective (cost)
in learning and generalizing the constitutive behavior of functions some criteria can be selected to be satisfied
complex materials and give better results (compared with through the search process. These criteria can be set in
conventional constitutive models) when they are employed order to (i) avoid the overfitting of models, (ii) push the
in a finite element code to analyse structures or domains models towards simpler structures, and (iii) avoid
made of the material under consideration. unnecessary terms representative of the noise in data. An
In this paper a new approach is introduced for interesting feature of EPR is in the possibility of getting
constitutive modeling of complex materials, which more than one model for a complex phenomenon. A further
integrates numerical and symbolic regression to perform feature of EPR is the high level of interactivity between the
evolutionary polynomial regression (EPR). The strategy user and the methodology. The user’s physical insight can
uses polynomial structures to take advantage of their be used to make hypotheses on the elements of the target
favorable mathematical properties. The main idea behind function and on its structure (Eq. (1)). Selecting an
the EPR is to use evolutionary search for exponents of appropriate objective function, assuming pre-selected
polynomial expressions by means of a genetic algorithm elements in Eq. (1) based on engineering judgment, and
(GA) engine. This allows (i) easy computational working with dimensional information enable refinement of
implementation of the algorithm, (ii) efficient search for an final models. Detailed explanation can be found in [18],
explicit expression (formula) and (iii) improved control of [19].
the complexity of the expression generated [18]. In what
follows, the main principles of EPR will be outlined. A III. CONSTITUTIVE MODELLING USING EPR
procedure is presented for computing the stiffness matrix In constitutive modeling using EPR, the raw experimental
using the trained EPR model and incorporation of the or in-situ test data are directly used for training the EPR. In
EPRCM in the finite element software ABAQUS. The this approach, there are no mathematical models to select
application of the developed EPR-based finite element and the EPR learns the constitutive relationships directly
 ISAST Transactions on Computers and Intelligent Systems, No. 2, Vol. 1, 2009
A. Javadi et.al: An Artificial Intelligence Based Finite Element Method
3

from the raw data during the training process. As a result, and then a forward pass through the neural network yields
there are no material parameters to be identified and as the prediction of the next expected state of stress and/or
more data becomes available, the material model can be strain relevant to an input strain or stress increment [7].
improved by re-training of the EPR using the additional In this paper, a method is introduced based on EPR for
data. Furthermore, the incorporation of an EPR in a finite constitutive modeling in finite element analysis. This
element procedure avoids the need for complex method takes advantage of the explicit mathematical
yielding/failure functions, flow rules, etc. An EPR equation representation of the relationships in EPR.
can be incorporated in a finite element code/procedure in To evaluate the potential of using EPR to derive
the same way as a conventional constitutive model. It can be functions describing the constitutive behavior of materials
incorporated either as incremental or total stress-strain two sets of stress-strain data are employed to train and test
strategies. In this study both the incremental and total stress- the EPR models. In both cases the data is divided into two
strain strategies have been successfully implemented in the separate sets. One set is used for training of the EPR model
intelligent finite element model. and the other one is used for validation to appraise the
generalization capability of the trained EPR model. After
A. Input and output parameters
training and validation, the best function is selected, based
The choice of input and output quantities is determined on the quality of fit according to the coefficient of
by both the source of the data and the way the trained EPR determination (COD) and also how well the selected model
model is to be used. A typical scheme to train most of the represents the actual stress-strain behavior.
neural network based material models includes an input set The first set of strain-stress data is representing a material
providing the network with the information relating to the with linear elastic behavior. The selected EPR model for the
current state units (e.g., current stresses and current strains) curve passing through data points is:

σ = 8.136 × 1013 ε 5 − 2.839 × 1011 ε 4 + 3.115 × 108 ε 3

250 (2)
− 8.285 × 10 4 ε 2 + 2.10 × 1011ε − 5.921 × 10−3
200 where ε is the strain and σ is corresponding stress.
The second set of data is corresponding to a material with
elasto-plastic behavior. After training EPR, Eq. 3 is selected
Stress (MPa)

150
as the best EPR model based on the COD.
100 Original Data
EPR
σ = 2.167 × 1011 ε − 6.297 × 1015 ε 3 +1.034 × 1018 ε 4
50 − 8.032 × 1019 ε 5 + 3.441 × 1021 ε 6 − 8.024 × 10 22 ε 7
(3)
+ 8.766 × 1023 ε 8 − 2.545 × 10 24 ε 9 − 6.505 × 106
0
0 0.0002 0.0004 0.0006 0.0008 0.001
Strain Figs. 1a and 1b show the stress-strain curve predicted by
Eqs. 2 and 3 (as a marker points) against those expected. It
(a)
1400
can be seen from these figures that EPR has successfully
captured the material behavior with an excellent accuracy.
1200 The material model in an FE analysis has to provide the
material stiffness matrix also known as the Jacobian. For
1000
infinitesimal strain increments ( dε ), J is the continuum
Stress (MPa)

800 Jacobian, J c
600 ∂ ( dσ )
Jc = (4)
Original Data ∂ ( dε )
400
EPR
This equation will be employed later to build the stiffness
200 matrix.
0
0 0.005 0.01 0.015 0.02 0.025 IV. INTELLIGENT FINITE ELEMENTS
Strain The obtained EPRCMs are implemented in a widely used
(b) general-purpose finite element code ABAQUS through the
user defined material module (UMAT). UMAT updates the
stresses and provides the material Jacobian matrix for every
Fig. 1. Results of the EPR models prediction and the original data.
increment at every integration point [20], [21]. In the
developed methodology (EPR–FEM), the EPRCM replaces
 ISAST Transactions on Computers and Intelligent Systems, No. 2, Vol. 1, 2009
A. Javadi et.al: An Artificial Intelligence Based Finite Element Method
4

the role of a conventional constitutive model. The source of incorporated in a FE code, is described in Fig. 2. This figure
knowledge for EPR is a set of raw experimental (or in situ) also shows the main steps of EPR.
data representing the mechanical response of the material to The constitutive relationship are generally given in the
applied load. When EPR is used for constitutive form [22]
description, the physical nature of the input–output data for
the EPR is clearly determined by the measured quantities, ∆σ = D∆ε (5)
e.g., stresses, strains, etc. The manner, in which EPRCM is

EPR FEA

Start

Input Data (Geometry,

Start Applied Load, Initial and
Boundary Condition)

Input Data (Experimental Increase the Applied Load

Data, Physical Insight) Incrementally

Current state of
stresses and strains
Genetic

Load Increment
Algorithm UMAT

Loop
EPRCM(s)

Mathematical 1- Next state of

Structure Stresses

2- Jacobian Matrix
Iteration Loop

Least Square
Solve the Main
Equation
EPR
Constitutive
equation
Symbolic NO
Function Convergence

YES

Fitness Output Result

Whole load NO
YES applied?
NO Check based on fitness criteria
and or generation number
YES

STOP

Fig. 2. The incorporation of EPR-based material model in ABAQUS finite element software for an integration point.
 ISAST Transactions on Computers and Intelligent Systems, No. 2, Vol. 1, 2009
A. Javadi et.al: An Artificial Intelligence Based Finite Element Method
5

where D is material stiffness matrix also known as the The EPR-based finite element model incorporating the
Jacobian. Assuming that matrix is elastic and isotropic for a trained EPR was used to analyse the behavior of the
load increment, matrix D is given in terms of Young’s cylinder under applied internal pressure. Assuming a linear
modulus, E, and Poisson’s ratio, ν. For a plane strain case, behavior for a small load increment in the nonlinear FE
for example analysis, the tangential elastic modulus of the material at
each strain can be obtained from the derivative of the Eq. 2.
1 − ν ν ν 0  Therefore the EPR based elastic modulus can be taken as:
 
 ν 1 − ν ν 0 
E dσ
D=  ν ν 1 −ν 0  (6) Et = = 4.068 × 1014 ε 4 − 1.1356 × 1012 ε 3
(1 + ν )(1 − 2ν )  dε (7)

 0 1 − 2ν  + 9.345 × 108 ε 2 − 1.657 × 105 ε + 2.1 × 1011
0 0
 2 
Eq. 7 is used to calculate stiffness matrix (Eq. 6). During
A. Numerical Examples the analysis, the Poisson’s ratio was kept constant.
1) Example 1: This example involves a thick circular
cylinder conforming to plane strain conditions. Fig. 3 shows
the geometric dimensions and the element discretization
employed in the solution and it is seen that 12 parabolic
isoparametric elements have been used. The cylinder is 200

made of linear elastic material with a Young's modulus of

E=2.1×105 N/mm2 and a Poisson's ratio of 0.3 [22]. This 160

example was deliberately kept simple in order to verify the Standard FEM
Stress (MPa)

IFEM (EPRCM)
computational methodology by comparing the results of a 120

linear elastic finite element model. The loading case

considered involves an internal pressure of 8.0×104 kN with 80

boundary conditions as shown in Fig. 3.

40

0
0 40 80 120 160 200
r (mm)

(a)
0.08
Radial Displacement (mm)

0.06

Standard FEM
IFE (EPRCM)
0.04

0.02
P
0
0 40 80 120 160 200
r (mm)
100 mm
(b)
200 mm

Fig. 4. Comparison of the results of the EPR-FEM and standard FEM in

Fig. 3. FE Mesh in symmetric quadrant of a thick cylinder. terms of
(a) radial stress and (b) radial displacement
 ISAST Transactions on Computers and Intelligent Systems, No. 2, Vol. 1, 2009
A. Javadi et.al: An Artificial Intelligence Based Finite Element Method
6

The results are compared with those obtained using a

standard linear elastic finite element method. Fig. 4 shows
the radial displacements and radial stresses along a radius of 50
the cylinder, predicted by the two different methods.
Comparison of the results shows that the results obtained

Applied Pressure (MPa)

40
using the EPR based FEM are in excellent agreement with
those attained from the standard finite element analysis. 30
This shows the potential of the developed EPR based finite
element method in deriving constitutive relationships from 20
raw data using EPR and using these relationships to solve Elasto_Plastic FE Analysis

boundary value problems. 10 EPRCM FE Analysis

2) Example 2: The second example is a plane stress beam 0

(Fig. 5) subjected to uniform pressure. The developed EPR 0 1 2 3

displacement (mm)
4 5

constitutive model (Eq. 3) is used to describe the material

behavior. To evaluate the described methodology, Fig. 6. Comparison of the results of the EPR-based FEM and
displacement of point A (mid-span) of the beam predicted conventional FEM
using the EPR-based FE analysis is compared with that
from conventional FE analysis. For the conventional FE V. CONCLUSIONS
analysis, an elasto-plastic model in ABAQUS, based on the An intelligent finite element method (EPR-FEM) has
tabulated stress-strain data, is used. been developed based on the integration of an EPRCM in a
For the EPR-based FE analysis, the Young’s modulus is finite element framework. In the developed methodology,
determined as: the EPRCM is used as an alternative to the conventional
constitutive models for the material. A procedure is
dσ presented for computing the stiffness matrix using the
Et = = 2.167 × 1011 − 1.889 × 1016 ε 2 + 4.137 × 1018 ε 3
dε trained EPR model and incorporation of the EPRCM in a
− 4.016 × 1020 ε 4 + 2.064 × 1022 ε 5 − 5.616 × 10 23 ε 6 (8)
commercial finite element code ABAQUS. The efficiency
+ 7.013 × 10 ε − 2.272 × 10 ε
24 7 25 8
and adaptability of the proposed method have been
demonstrated by successful application to two boundary
Fig. 6 shows the load-displacement curves at point A value problems. The results of the analysis have been
(mid-span of the beam) obtained using the conventional FE compared to those obtained from conventional FE analyses
model and the developed EPR-based FE model. It is shown using the linear elastic and elastic-plastic models. The result
that results of the EPR-based FEM are in very good shows that EPRCM can be successfully implemented in a
agreement with those from the conventional FE analysis finite element model as an effective alternative to
using an elasto-plastic model. conventional material models. It is also shown that stiffness
matrix elements can be directly obtained from EPR stress-
strain relationship.

REFERENCES
[1] H.S. Shin, Neural network based constitutive models for finite
element analysis, PhD Dissertation, University of Wales, Swansea,
100 mm UK, 2001.
[2] J. Ghaboussi, J. Carret and X. Wu, Knowledge-based modeling of
material behavior with neural networks. Journal of Engineering
A Mechanics Division, vol. 17, pp. 32-153, 1991.
[3] J. Ghaboussi and D.E. Sidarta, New nested adaptive neural networks
600 mm (NANN) for constitutive modeling. Computers and Geotechnics, vol.
22, pp. 29-52, 1998.
Fig. 5. Simply supported beam under uniform pressure [4] D. Penumadu and R. Zhao, Triaxial compression behavior of sand
and gravel using artificial neural networks. Computers and
Geotechnics, vol. 24, pp. 207-230, 1999.
[5] G.W. Ellis, C. Yao, R. Zhao, D. Penumadu, Stress–strain modeling
of sands using artificial neural netwoks, ASCE Journal of
Geotechnical Engineering Division vol. 121, pp. 429–435, 1995.
[6] J.-H. Zhu, M.M. Zaman, S.A. Anderson, Modeling of soil behavior
with a recurrent neural network, Canadian Geotechnical Journal vol.
35, pp. 858–872, 1998.
[7] J. Ghaboussi, D.A. Pecknold, M. Zhang, R.M. Haj-Ali,
Autoprogressive training of neural network constitutive models,
 ISAST Transactions on Computers and Intelligent Systems, No. 2, Vol. 1, 2009
A. Javadi et.al: An Artificial Intelligence Based Finite Element Method
7

International Journal for Numerical Methods in Engineering vol. 42,

pp. 105–126, 1998.
[8] D.E. Sidarta, J. Ghaboussi, Constitutive modeling of geomaterials
from nonuniform material tests, Computers and Geotechnics vol.
22,pp. 53–71,1998.
[9] D. Penumadu, J.L. Chameau, Geomaterial modeling using neural
networks, in: N. Kartman, I. Flood, J.H. Garrett (Eds.), Artificial
Neural Networks for Civil Engineering: Fundamentals and
Applications, ASCE, 1997, pp. 160–184.
[10] A.A. Javadi and M. Rezania. Applications of artificial intelligence
and data mining techniques in soil modeling. Geomechanics and
Engineering, an International Journal, vol. 1, pp. 53-74, 2009.
[11] H.S. Shin, G.N. Pande, Enhancement of data for training neural
network based constitutive models for geomaterials, in: Proceedings
of the Eighth International Symposium on Numerical Models in
Geomechanics-NUMOG VIII, Rome, Italy, pp. 141–146, 2002.
[12] H.S. Shin and G.N. Pande, On self-learning finite element code
based on monitored response of structures. Computers and
Geotechnics, vol. 27, pp. 161-178, 2000.
[13] Y.M. Hashash, S. Jung and J. Ghaboussi, Numerical implementation
of a neural network based material model in finite element analysis.
Int. Journal for Numerical Methods in Engineering, vol. 5, pp. 989-
1005 2004.
[14] A.A. Javadi, T.P. Tan and M. Zhang, Neural network for constitutive
modeling in finite element analysis. Computer Assisted Mechanics
and Engineering Sciences, vol. 10, pp. 375-381, 2003.
[15] A.A. Javadi, M. Zhang, T.P. Tan, Neural network for constitutive
modeling of material in finite element analysis, in: Proceedings of
the Third International Workshop/Euroconference on Trefftz
Method, Exeter, UK, pp. 61–62, 2002.
[16] A.A. Javadi, T.P. Tan, A.S.I. Elkassas, An intelligent finite element
method, in: Proceedings of the 11th International EG-ICE
Workshop, Weimar, Germany, pp. 16–25, 2004.
[17] A.A. Javadi, T.P. Tan, A.S.I. Elkassas, Intelligent finite element
method, in: Proceedings of the Third MIT Conference on
Computational Fluid and Solid Mechanics, Cambridge,
Massachusetts, USA, pp. 347–350, 2005.
[18] O. Giustolisi and D.A. Savic. A Symbolic Data-driven Technique
Based on Evolutionary Polynomial Regression. Journal of
Hydroinformatics, vol. 8, pp. 207–222, 2006.
[19] A. Doglioni, O. Giustolisi, D.A. Savic, B.W. Webb, An investigation
on stream temperature analysis based on evolutionary computing,
Hydrological Processes vol. 22, pp. 315–326, 2008.
[20] ABAQUS User Subroutines Reference Manual (version 6.7-1)
Dassault Systems, 2007.
[21] ABAQUS Theory Manual (version 6.7-1) Dassault Systems, 2007.
[22] D.R.J. Owen, E. Hinton, Finite Elements in Plasticity: Theory and
Practice, Pineridge Press, Swansea, 1980.

View publication stats