Mid Sem Report
Mid Sem Report
Mid Sem Report
ID Numbers
S Ananth Narayan
2009AAPS117H
G Sharath Chandra
2009A7PS149H
By
B.Aditya
Project Guide
Dr. Amit Kumar Gupta, Assistant Professor, Mechanical Engineering Department
March 2012
CERTIFICATE
This is to certify that the report entitled Application of ANN, DE and GP in Deep
Drawing operations, and submitted by B. Aditya, ID No. 2010A4PS155H in partial
fulfillment of the requirements of BITS C331 Computer Projects embodies the work done
by him under my supervision.
Page | 0
ACKNOWLEDGEMENTS
I express my gratitude to Dr. Amit Kumar Gupta for giving me the ideas, suggestions to work in this
field. He has been giving me motivation and also an opportunity to work with him and gain knowledge
in the field of research.
I would like to acknowledge the efforts of the faculty and director of BITS Pilani Hyderabad Campus
for providing us with such a course where we can go beyond the normal course work and explore the
current developments in various fields.
Page | 1
ABSTRACT
Application of soft computing techniques in the field of mechanical engineering is the gap to be filled by
the researchers. Now days, computers are used in every corner of our lives. The integration of soft
computing techniques is the key to obtain good results in terms of more accuracy, and extrapolate the
controllability by optimizing the parameters.
In this report we shall first see the application of the soft computing techniques in the field of material
science. The Flow stress prediction using constitutive models is one of the areas of research under
consideration for many researchers. Its importance comes from its usage in Finite Element Modeling as
material models. Intelligent FEM systems are an area which has not been studied properly by many
researchers. This project makes some study on the implementation of the intelligent material Models in
FEM simulations. This will improve the prediction offered by the FEM simulation to the experimental
values.
Page | 2
Contents
ACKNOWLEDGEMENTS ................................................................................................................. 0
CERTIFICATE............................................................................................................................ 0
ACKNOWLEDGEMENTS ......................................................................................................... 1
ABSTRACT ................................................................................................................................. 2
FLOW STRESS PREDICTION USING SOFT COMPUTING TECHNIQUES FOR TI-6AL-4V ALLOYS. ... 4
Introduction....................................................................................................................................... 4
Materials and experimental details ........................................................................................... 6
Models for prediction of flow stress ........................................................................................... 7
ANN Model ............................................................................................................................ 7
SVM Models ........................................................................................................................... 7
Implementation of the Intelligent models ................................................................................... 9
ANN fine tuned with Differential Evolution Algorithm.................................................... 9
LSSVM ................................................................................................................................... 11
SVM-KM ................................................................................................................................ 11
Result and discussion .................................................................................................................... 11
ANN model predictions ...................................................................................................... 11
LS-SVM model predictions................................................................................................. 11
SVM-KM model predictions............................................................................................... 11
Comparison of ANN, LS-SVM, SVM-KM model predictionsError! Bookmark not defined.
Conclusions ..................................................................................................................................... 12
References:..................................................................................................................................... 13
Page | 3
FLOW STRESS PREDICTION USING SOFT COMPUTING TECHNIQUES FOR TI6AL-4V ALLOYS.
The various models for predicting the Flow stress at elevated temperatures have been studied. Models
such as ANN, LSSVM, SVM-KM, and ELM are used for predicting the flow stress of the Ti-6Al-4V
alloy. The tensile test experiments are conducted on Titanium Alloy at the temperatures of 500C-7000C
at intervals of 500C and at strain rates of 0.01, 0.001, 0.0001 and 0.00001s-1. The test results are
compared based on the computation time, pearsons correlation coefficient and the testing accuracy. The
goodness of fit of the model is evaluated based on the statistical tests of t-test for means and f-test and
levenes test for variance in the predictions. The tests suggest that all the models have a good fitness
with the experiments.
Introduction
Ti-6Al-4V alloy presently is the most broadly used alloy, accounting for more than 50% of all titanium
tonnage in the world [1]. It is designed primarily for high strength at low to moderate temperatures; it
has a high specific strength (strength/ density), stability at temperature up to 4000C, and good corrosion
resistance [2]. Ti-6Al-4V alloy is used comprehensively for turbine engines, airframe, engine
components include blades, discs, wheels. In addition the superplastic characteristics of fine grained,
equiaxed Ti-6Al-4V alloy is being used progressively for aerospace applications. It also has good
diffusion-bonding characteristics, which combined with superplastic forming, enables the fabrication of
very complex structures [1-2]. Its outstanding corrosion/erosion resistance provides the prime
motivation for chemical process, marine use, food processing plants, oil refinery heat exchangers,
medical prostheses and industrial use. Currently, titanium alloys are common, readily available
engineered metals that compete directly with stainless and specialty steels, copper alloys, nickel based
alloys and composites [3-4]. This wide spectrum of application, most of them being very critical with
very consequences, necessitates a better understanding of material behavior of the alloy at various
temperatures and loading conditions.
A constitutive model defines the relationship of the dynamic material properties with process
parameters. The major influencing parameters are strain, strain rate and deformation temperature [5-6].
Conventionally, the methods of regression analysis and statistic were carried out to obtain the material
constants for developing constitutive models [7]. The complex interconnections between these
parameters and materials properties, mathematical models are sometimes very complex to handle by the
numerical techniques as well as by experimental methods particularly when it includes some particular
material phenomenon such as dynamic strain ageing (DSA). DSA is considered by jagged stressstrain
Page | 4
curve, i.e., wavy pattern on stressstrain curve. DSA is also known as Portevin-Le Chatelier (PLC)
effect [6]. Considerable amount of the works were done in past few decades to link the flow stress with
the process parameters through the constitutive and the empirical models [8]. The finite element-based
models were also described to predict the flow stress which requires definite initial and boundary
conditions. Furthermore, the clear identifications of the boundary conditions are difficult; particularly
the measurements of the coefficient of friction and heat transfer at the interface [9-10]. Ti-6Al-4V alloy
is the typical two-phase (+) alloy. As an alpha beta alloy, Ti-6Al-4V may have different volume
fraction of alpha and beta phases, depending on heat treatment and interstitial content [1-4]. As of now,
studies in the field of constitutive modeling have been mostly focused on FCC and BCC metals due to
their simple crystalline structures. Slight attention has been done on the hexagonal close-packed (HCP)
metals, particularly alloys of a more complicated crystalline structure, such as Ti-6Al-4V alloy [11].
Very few constitutive models were available for two phase Ti-6Al-4V alloy [11-15].
Nowadays, intelligent theories and methods have been used for prediction, estimation and optimization
of regression problems such as artificial neural network (ANN), support vector machines (SVM). The
emergence of intelligent theories and methods has been provided many promising results in the field of
materials science simulations. Advantage of these methods are one does not need to have a well-defined
process for algorithmically converting an input to an output. Rather, it requires only a collection of
representative examples of the desired mapping. The application of soft computing techniques are the
novel way to study the high temperature deformation behavior and few efforts have been made to the
applications of ANN in some alloys [6-7, 17-22]. Concern to titanium alloy very few literatures have
been reported. Li et al. [17] established the predicting model for the calculation of flow stress of Ti-153
alloy based on the ANN method. Reddy et al. [16, 17] developed a back propagation neural network
model to predict the flow stress of Ti6Al4V alloy over a temperature range of 700011000C and at the
strain rates of 104 to 102 s1. Recently, support vector machines (SVM) has been introduced to solve
machine learning tasks such as pattern recognition, regression and estimation. Due to its excellent
properties of globally optimal solution and good learning ability, SVM attracted wide attention. Yan Lou
et al. [23] attempted to develop high temperature flow stress of AZ80 Magnesium Alloy with particle
swarm optimization-based support vector regression. Datasets were established based on compression
tests in the temperature range of 3500-4500C and strain rate range of 0.0150s1.
Page | 5
However, there is hardly any reported work on comparative analysis of the intelligent methods to predict
flow stress in DSA and non-DSA region of commercially important two-phase Ti-6Al-4V alloy. In the
present investigation, efforts have been made to develop a model to predict flow stress using ANN,
LSSVM, SVM-KM and ELM methods and experimental verification of the same.
(dark phase). The as-received initial microstructure was laminar contained about 90% of phase and
about 10% of phase, as shown in Figure 2. The average grain size of both phase and the transformed
phase were 6-8 m. A slower cooling rate (e.g., air or furnace cooling) permits diffusion-controlled
partitioning between and as the temperature falls below the transus and the nucleation of begins.
The microstructure is laminar, in which broad and fine lamellae alternate to form packets. The large
volume fraction of the alpha grains was predominantly non-uniform in size and shape. The net volume
fraction of the beta phase was low when compared to the alpha phase.
Artificial neural network is authoritative data information treatment system which attempts to simulate
the neural networks structure of the human brain. It can signify and capture complex non-linear
interactions between inputs and outputs. Each neural network is composed of an input layer, an output
layer and one or more hidden layers, which are connected by the processing units called neurons. Each
neuron works as a self-governing processing element, and has a related transfer function, which defines
how the weighted sum of its inputs is transformed to the results into an output value. Among the various
kinds of ANN approaches, the back propagation (BP) learning algorithm has become the most popular
in engineering applications. Back propagation algorithm is based on minimization of the quadratic cost
function by tuning the network parameters. The mean square error (MSE) is considered as a
measurement criterion for a training set. Specially, BP neural network is the most suitable tool for
treating non-linear systems.
SVM Models
A support vector machine (SVM) is a part of supervised learning, a branch of statistical learning which
acquires through a series of examples and becomes trained, i.e., it creates a decision-maker system
which attempts to forecast new values. When SVM is applied to regression problems, then it is called
support vector regression (SVR). ANNs and SVRs show comparable results on the most popular
benchmark problems, the theoretical status of SVRs makes them an attractive and promising area of
research [24]. The SVR was presented as a learning technique that initiated from the theoretical
foundations of statistical learning theory and structural risk minimization. The SVR first non-linearly
transforms the original input space x into a higher dimensional feature space. That is, in order to learn
non-linear relations with a linear machine, it is essential to select a set of non-linear features and to
Page | 7
express the data in the new representation. This is equivalent to applying a fixed non-linear mapping of
the data to a feature space which the linear machine can be used in. This transformation can be attained
by using various types of non-linear mapping. Non-linear regression problems in an input space can
become linear regression problems in a feature space.
N
The SVR model is given N training data ( xi , y i ) R m R where xi is the input vector to
i 1
the SVR model and yi is the actual output value, from which it learns the input output relationship. The
SVR model can be represented as follows [25].
N
y f ( x) i i ( x) b T ( x) b
i 1
Where the function i(x) is called the feature that is non-linearly mapped from the input space x,
Page | 8
Least Squares Support Vector Machines (LS-SVMs) are reformulations to standard SVMs. LS-SVMs
are closely related to regularization networks and Gaussian processes but additionally emphasize and
exploit primal-dual interpretations from optimization theory. The Formuation of LS-SVM is similar to
regular SVM. The reader is suggested to refer to (Suykens et al. 2002), for further reference.
SVM-KM MODEL
The training vectors near the separation margins during the optimization phase of SVM are likely to
become support vector and must be preserved. Whereas the training vectors far from the margins are not
in general taken into account for the SVM's design process. The SVM-KM toolbox groups the training
vectors in many clusters. Consequently, the number of vectors in SVM-KM training is smaller and the
training time can be decreased without compromising the generalization capability of the SVM. The
reader is suggested to refer to (Barros de Almeida, M.et al) for better understanding of the formulation
of SVM-KM.
A back propagation algorithm was used to train a feed forward neural network, which is consistent and
most commonly utilized. In this study, the input variables of ANN include strain, strain rate and
deformation temperature, whereas the output variable is flow stress. Feed forward network trained with
the back propagation algorithm was developed, as shown in Figure 3. Before training the network, the
input and output datasets have been normalized within the range of 0.050.95 to avert a specific factor
from dominating the learning for the ANN model. Thus, using Eq. (1), the experimental data was
normalized to make the neural network training more effective prior to the use of the datasets.
(1)
Architecture selection requires choosing the appropriate number of hidden units. In most of the
applications of ANN, this selection was done using a trial-and-error procedure. The number of hidden
layers defines the complexity of neural network and precision of predicted values. If the architecture is
too complex, it may not converge during training or the trained data might be over fitted. In other way,
the trained network may not have adequate capability to acquire the process appropriately. Therefore,
several networks with varying number of neurons in hidden layer were observed. The value of mean
Page | 9
square error (MSE) is used to check the capability of a particular architecture. Differential Evolution
algorithm (DE) was used for training the networks and for the tuning of weights so as to determine the
optimal architecture of neural network.
The DE algorithm gives the optimal architecture after evaluating objectives.
Like all evolutionary algorithms, DE maintains a population and evolves to better solutions to yield the
optimal settings of the problem. At the beginning all members of the population are initialized
randomly. The algorithm advances in generations. Each generation involves three consecutive phases:
reproduction (creation of a temporary population), computing of the objective function (called the
fitness in the EA terminology) for each temporary population member, and selection. Reproduction in
differential evolution creates a temporary population of trial vectors. For each solution a corresponding
trial vector is created. Each element of the trial vector is generated as:
If r<CR
Elsewhere
(2)
In the above expression, value of F is between 0 and 2. It is a user supplied parameter called the
mutation coefficient. The parameters F (is called the differential weight) and CR (is called the crossover
probability) influence the convergence speed and robustness of the optimization process. In this problem
the value of F and CR are 0.8 and 0.7 respectively which are taken from literature.
The remaining two phases of a single DE generation are the computation of fitness for all members of
the trial population and the selection. The selection in differential evolution is very simple. The fitness
of each trial solution is compared to the fitness of the corresponding original solution. The trial vector
replaces the original if its fitness is better. Otherwise the trial vector is discarded.
To apply DE to ANN training, the algorithm is used to minimize the sum of squared errors (SSE). The
evaluation of this function requires iterating through all elements of the training set and summing all the
partial results (squared errors in the case of SSE) obtained for all the elements. But the SSE since the
testing and training phase play and important role in the neural network performance, the objective
function of the DE algorithm is used as a linear combination of the SSE of prediction in both training
phase and SSE of prediction in testing phase.
Page | 10
Once the best trained network is found, all the transformed data returns to their original value using the
following equation:
(4)
LSSVM
LSSVM is modelled in a similar way to that of ANN. The implementation is done using the LSSVM
toolbox based on MATLAB environment. First the data is normalised between 0.05 and 0.95 using
equation (1). The parameters which specify the margin of the hyper plane in the feature space are gam
and sig in the implementation provided by Canu et. al. in the toolbox. The optimization of these values
gives a good training over the data. These parameters are tuned using DE algorithm with SSE as the
objective function which is to be minimised. Once the best set of gam and sig are evaluated, the
predictions are transformed using equation (4).
SVM-KM
The implementation of the SVM-KM is completely similar to that of LSSVM. Fine-tuning of parameters
was again done with the help of DE algorithm. The preprocessing of the data and the post-processing
remains same for all the models discussed.
Conclusions
The different models for the prediction of Flow Stress as a function of strain, strain rate and temperature were
developed. The study of many more models is the future of the work to be done.
Page | 12
References:
1. Joseph D. Beal, Rodney Boyer, and Daniel Sanders, The Boeing Company, Forming of
Titanium and Titanium Alloys, ASM Handbook, Volume 14B: Metalworking: Sheet Forming,
pp.656-669.
2. Rodney Boyer, Gerhard Welsch, E.W.Collings, Materials Properties Handbook: Titanium
Alloys, ASM International, ISBN 0-87170-481-1, pp 483-484.
3. MJ Donachie. Titanium-A Technical Guide, American Society for Metals, Metals Park, OH,
1988.
4. Edited by, C. Leyens and M. Peters, Titanium and Titanium alloys, Wiley Publications.
5. Guo Z, Sha W. Modelling the correlation between processing parameters and properties of
maraging steels using artificial neural network. Comput Mater Sci. 2004; 29:1228.
6. Amit Kumar Gupta, Swadesh Kumar Singh , Swathi Reddy , Gokul Hariharan , Prediction of
flow stress in dynamic strain aging regime of austenitic stainless steel 316 using artificial neural
network. Materials and Design 35 (2012) 589595.
7. Y. Sun , W.D. Zeng , Y.Q. Zhao , X.M. Zhang , Y. Shu , Y.G. Zhou, Modeling constitutive
relationship of Ti40 alloy using artificial neural network, Materials and Design 32 (2011) 1537
1541.
8. Y.C. Lin , Xiao-Min Chen, A critical review of experimental results and constitutive
descriptions for metals and alloys in hot working, Materials and Design 32 (2011) 17331759.
9. Carlos Garca-Garino, Felipe Gabaldnb, Jos M. Goicolea, Finite element simulation of the
simple tension test in metals, Finite Elements in Analysis and Design 42 (2006) 1187 1197.
10. A. Arriagaa, J.M. Lazkano, R. Pagaldai, A.M. Zaldua, R. Hernandez, R. Atxurra, A.
Chrysostomou, Finite-element analysis of quasi-static characterisation tests in thermoplastic
materials: Experimental and numerical analysis results correlation with ANSYS, Polymer
Testing 26 (2007) 284305.
11. Woei-Shyan Lee, Chi-Feng Lin, Plastic deformation and fracture behaviour of Ti6Al4V alloy
loaded with high strain rate under various temperatures, Materials Science and Engineering
A241 (1998) 4859.
12. S. Nemat-Nasser, W.G. Guo, V.F. Nesterenko, S.S. Infrakanti, Y.B. Gu, Dynamic response of
conventional and hot isostatically pressed Ti-6Al-4V alloy: experiments and modeling, Mech.
Mater. 33 (2001) 425439.
13. R.C.Picu, A. Majorell, Mechanical behavior of Ti6Al4V at high and moderate temperaturesPart II: Constitutive modeling, Material Science and Engineering A326 (2002), 306-316.
14. A.S. Khan, R. Kazmi, B. Farrokh, M. Zupan, Multiaxial and non-proportional loading
responses, anisotropy and modeling of Ti6Al4V titanium alloy over wide ranges of strain rates
and temperatures, Int. J. Plasticity 23 (2007) 11051125.
15. C.Y. Gao, L.C. Zhang, H.X. Yana, A new constitutive model for HCP metals, Materials
Science and Engineering A 528 (2011) 44454452.
Page | 13
16. Jun Cai , Fuguo Li, Taiying Liu, Bo Chen, Min He, Constitutive equations for elevated
temperature flow stress of Ti6Al4V alloy considering the effect of strain, Materials and
Design 32 (2011) 11441151.
17. Li Ping , Xue Kemin , Lu Yan, Tan Jianrong, Neural network prediction of flow stress of Ti
153 alloy under hot compression, Journal of Materials Processing Technology 148 (2004)
235238.
18. Madakasira Prabhakar Phaniraj, Ashok Kumar Lahiri, The applicability of neural network
model to predict flow stress for carbon steels, Journal of Materials Processing Technology 141
(2003) 219227.
19. Y.C. Lin a, Jun Zhang , Jue Zhong, Application of neural networks to predict the elevated
temperature flow behavior of a low alloy steel, Computational Materials Science 43 (2008)
752758.
20. Sumantra Mandal , P.V. Sivaprasad , S. Venugopal , K.P.N. Murthy, Artificial neural network
modeling to evaluate and predict the deformation behavior of stainless steel type AISI 304L
during hot torsion, Applied Soft Computing 9 (2009) 237244.
21. Yanchun Zhu, Weidong Zeng , Yu Sun, Fei Feng, Yigang Zhou, Artificial neural network
approach to predict the flow stress in the isothermal compression of as-cast TC21 titanium
alloy, Computational Materials Science 50 (2011) 17851790.
22. Guoliang Ji, Fuguo Li , Qinghua Li , Huiqu Li , Zhi Li, Prediction of the hot deformation
behavior for Aermet100 steel using an artificial neural network, Computational Materials
Science 48 (2010) 626632.
23. Yan Lou, ChangXing Ke ,Luoxing Li, Accurately Predicting High Temperature Flow Stress of
AZ80 Magnesium Alloy with Particle Swarm Optimization-based Support Vector Regression,
Appl. Math. Inf. Sci. 7, No. 3, 2013, 1093-1102.
24. Amit Kumar Gupta, Predictive modelling of turning operations using response surface
methodology, artificial neural networks and support vector regression, International Journal of
Production Research, Vol. 48, No. 3, 1 February 2010, 763778.
25. Parrella, F., 2007. Online support vector regression. A thesis presented for the degree of
Information Science, University of Genoa, Italy.
Page | 14
List of Figures:
Figure 1: Flat tensile test specimens of ASS 316 with 1 mm thickness as per ASTM sub-size Standards
Figure 2 (a): A computer controlled UTM with a high temperature chamber.
Figure 2(b): Resistance heating Split Furnace
Figure 3(a): Optical photograph of as-received microstructure of Ti6Al4V alloy
Figure 3 (b) & (c) indicates microstructural observation of non DSA and DSA region samples.
Figure 4. Schematic layout of the neural network architecture.
Figure 5. Impact of hidden neutrons on the network performance.
Figure 6. Prediction of ANN model.
Page | 15
Page | 17
AL
Fe
Ti
Comp%
5.56
4.07
0.185
0.022
89.997
Page | 19
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
No
Yes
Yes
Yes
Yes
Note: Attention should be paid on the above items while preparing the FINAL REPORT. Supervisors
are also requested to ensure that their student(s) have prepared the FINAL REPORT properly.
The final project report shall be submitted to your supervisor. An electronic copy (single file in .pdf
format) of the final report is required to be submitted through your supervisor to Instructor-in-charge.
Page | 20