Materials and Design 36 (2012) 331–336

Contents lists available at SciVerse ScienceDirect

Materials and Design

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

Technical Report

Prediction of springback in sheet metal components with holes on the bending area,
using experiments, finite element and neural networks
V. Nasrollahi ⇑, B. Arezoo
CAD/CAM Research Center, Mechanical Engineering Department, Amirkabir University of Technology (Tehran Polytechnic), Hafez Street, Tehran, Iran

a r t i c l e i n f o a b s t r a c t

Article history: The springback phenomenon has a significant role in precision sheet metal bending. Traditionally, the
Received 28 August 2011 designers obtain the values for springback for different materials and bending parameters from handbook
Accepted 16 November 2011 tables or springback graphs. However these tables or graphs cannot be used for perforated components
Available online 25 November 2011
which are parts with holes on the bending surfaces. This paper presents the results of research in wipe
bending for perforated components. In the present research the influence of process variables such as
hole type, number of holes, the ratio of hole width to sheet width, die radius and pad force on springback
are discussed. Some experiments are carried out on HSLA360 and St12 materials which show that the
presence of holes on the bending area can affect the springback considerably. These experiments are also
simulated by finite element method. The results of the finite element model used, compared with exper-
iments show the reliability of the proposed model. These results are also used as the training data for two
artificial neural networks. The first network is used for one type of hole and the second one for three
types of holes. After testing both networks the results show that the latter is more accurate to predict
the springback.
Ó 2011 Elsevier Ltd. All rights reserved.

1. Introduction elastic–plastic pure bending of a rectangular plate. Leu [4] sug-

gested an equation to evaluate the bendability and springback in
Springback is an important issue in sheet metal forming and can plastic bending of anisotropic sheet metals using anisotropic Hill’s
lead to significant problems during component assembly if the theory. Buranathiti and Cao [5] calculated the bending moment in
phenomenon is not well controlled. It arises from the elastic recov- a straight flanging process and then predicted the springback pro-
ery, mainly due to bending. The concerns about springback and file. Zhang et al. [6] used their model to predict the springback in
quality control grow among auto-makers as the use of high U-shaped channels bent under plane-strain conditions. By using
strength materials increases. Parts made of high strength steels strains after relief from the maximum bending stress, Yi et al. [7]
demonstrate more springback than parts made of mild steel. also developed an analytical model for six different deformation
Factors which affect springback can be divided into two catego- patterns. They estimated springback after elastic recovery by the
ries. The factors which are related to material, like young’s modu- residual differential strains between outer and inner surfaces. Ana-
lus, yield stress, strength coefficient, strain hardening, poisson’s lytical models generally have to be formulated with significant
ratio and anisotropic coefficient. The other factors are those which simplifications of process conditions and material properties
are related to the bending process like tooling geometry, clearance, which can considerably affect the accuracy of the results.
pad force etc. Several papers have been published dealing with With the advent of computation technology, prediction and
these parameters and their effect on springback. In these works simulation of springback can be analyzed using finite element
there are three major approaches in analyzing and predicting methods. Prior [8], and Finn et al. [9], used explicit–implicit solvers
springback: the analytical method, the finite element analysis to simulate bending processes and the spring-back effect. Many
and artificial intelligence technologies such as neural networks. more articles have been published and focused on the effect of
The analytical approach is based upon simplifications of process bending parameters on springback using FEM. Esat et al. [10] con-
and material behavior. In an early attempt Gardiner [1] proposed sidered the effect of different aluminum materials of different
an analytical solution for springback for plane stress pure bending thicknesses. Chan et al. [11] investigated on punch radius, punch
with perfect elastoplasticity. Johnson and Yu [2,3] followed the angle and die-lip radius. Ling et al. [12] studied the effect of differ-
work of Gardiner to give an analysis of springback in the bi-axial, ent die parameters like die clearance, die radius, step height and
step distance for L-bending using FEM. Panthi et al. [13] proposed
⇑ Corresponding author. Tel.: +98 21 64 54 3472; fax: +98 21 6641 9736. indigenous finite element software to predict the springback and
E-mail address: [email protected] (V. Nasrollahi). asses the influence of bending parameters on springback.

0261-3069/$ - see front matter Ó 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.matdes.2011.11.039
332 V. Nasrollahi, B. Arezoo / Materials and Design 36 (2012) 331–336

Although FEMs are an useful tool to predict the springback, they The bending sequence and the tools are shown in Fig 2 for a part
are a time consuming processes. By training and using neural net- with three holes.
works this problem can be solved. Inamdar et al. [14] developed an The parameters and their different levels are specified in Table 1.
artificial network to predict springback in air V bending. Six As it is shown, three types of holes are considered (oblong, circle,
parameters (the die gap, the punch radius/thickness ratio, the die square) and they are positioned in the center of the bending area.
entry radius, the angle of bend, the yield strength to Young’s mod- The bending line is transverse to rolling direction.
ulus ratio, and the strain hardening exponent, n) were used as in- Another parameter is the ratio of hole width to sheet width
put to the network. Liu et al. [15] proposed a neural network (hole ratio). The Hole ratio is W/L as it is shown in Fig 3. When hole
prediction model which used genetic algorithm to optimize the ratio is equal to zero it means there are no holes in the bending
weights of neural networks. Kazan et al. [16] used FEM data to area.
train neural networks and determine springback in wipe bending.
As stated several authors studied the effect of bending parame-
3. Finite element simulation
ters and material on springback. However the effect of presence of
hole on springback was not widely considered in literatures. Farsi
Finite element analysis was carried out using the commercial
et al. [17] carried out an experimental investigation on the effect of
code Abaqus/Standard. Because of the hole, the FEM model has
holes on springback and bending forces in V bending. They consid-
to be 3D model. Fig 4 shows the set-up of the FEM model. Note that
ered different thicknesses, die widths and the ratio of hole width to
the sheet, loading and boundary conditions are symmetric, so only
sheet width. In this paper a more comprehensive study is carried
half of the geometry is used for simulation.
out and some more parameters are considered. The parameters
In metal forming simulation, one of the most important factors
used are hole type, number of holes, the ratio of hole width to sheet
is the relationship between stress and strain. The wipe bending set
width, die radius and pad force. Over 120 experiments were carried
is composed of the die, punch, pad and sheet. The pad, die, and
out on HSLA360 and St12 materials. These experiments are used
punch are defined as solid material, and the sheet metal is defined
for validating a FEM model which generates data for training a
as elastic–plastic.
neural network.
Duplicate ASTM E8M longitudinal tensile samples [19] were
tested for each sheet metal (HSLA360 & St12) on a Zwick universal
2. Experimental procedure testing machine. The result of tensile test is shown in Fig 5. By
using these curves the material behaviors are defined for the
Wipe bending is a common process in sheet metal forming. The FEM model. To get more precise model of sheet material, based
recommended parameters and dimensions for this process are on ASTM E517 (specimen A) [20], two tests are carried out on each
shown in Fig. 1. It should be noted that the values used in this fig- three directions (0°, 45° and 90° off rolling direction) to specify the
ure are from Ref. [18]. anisotropic behavior of material. The plastic ratio R is the ratio of
In this study a wipe bending apparatus with changeable die the true strain which occurs in a width direction perpendicular
radius is implemented. The pad force is also adjustable. to the direction of applied stress in the plane of the sheet, to the
The bending tests comprised the following steps. concomitant true strain in the thickness direction. R0, R90 and R45
are the R values when the direction of applied stress is along rolling
1. Checking the clearance and bending angle directions (0°), transverse direction (90°) and diagonal to the roll-
2. Replacing the part on the die and applying the supposed pad ing direction (45°) respectively. Rm is weighted average of R values
force obtained in three directions [20]:
3. Moving the punch down at a constant speed and bending the
component. R0 þ R90 þ 2R45
Rm ¼ ð1Þ
4. Moving the punch up. 4
5. Replacing the part, measuring its angle by a profile projector. The anisotropic and tensile test results plus the coefficients of
(Accuracy 1 min) stress–strain relationship in the plastic range (Eq. (2)) are listed
in Table 2.

rtrue ¼ K entrue ð2Þ

The meshing of the model is carried out by means of Hex ele-
ment. An 8-node linear brick is used (C3D8). Fine mesh was
applied in the bending zone of the specimen and the central hole
at which high local plastic flow will occur. To get the optimum ele-
ment size, element refinement is required to achieve convergence
of spring-back angles. The appropriate element size is chosen
based on the accuracy of spring-back angles and the time taken
for the simulation to complete.
The objectives of this analysis are to investigate the possibility
of predicting springback for sheet metal components with holes
on the bending area. Also this analysis is used to provide training
data for neural networks.

4. Architecture, training and testing the artificial Neural

Networks

The artificial neural networks (ANN) attempts to imitate the

learning activities of the human brain. Neural networks provide a
Fig. 1. Wipe bending parameters and the recommended dimensions. means to model the input/output relationship of data in a
 V. Nasrollahi, B. Arezoo / Materials and Design 36 (2012) 331–336 333

Fig. 2. Wipe bending sequences (a) initial position, (b) applying pad force and (c) bending.

Table 1
The parameters and their levels.

1st Level 2nd Level 3rd Level 4th Level 5th Level 6th Level
Hole type Oblong Circle Square – – –
Hole ratio 0.5 0 0.3 0.4 0.6 0.7
Die radius 4t 3t 2t – – –
Fpad/L (N/mm) 100 80 60 40 20 –
Material HSLA360 St 12 – – – –
Number of holes 1 2 3 – – –

Fig. 3. The definition of hole ratio W/L.

Fig. 5. The results of tensile tests based on ASTM E8 M for HSLA360 and St12.

layers between the input and output layers are called hidden layers
and consist of many computational neurons and transfer functions.
Usually these networks are fully connected which means that each
neuron in any layer of the network is connected to all neurons of
the previous layer. Back propagation is a gradient descent algorithm.
The difference between the network output obtained and the
desired output is compared and iterated again until the output
reaches the prescribed tolerance value. A complete description of
this algorithm can be found in [21] and [22].
Choosing the optimum architecture of the network is one of the
Fig. 4. The set-up of the FEM model.
challenging steps in neural network modeling. Unfortunately, this
step is usually a trial and error procedure and various situations
are required to be attempted [23]. So several fully connected ANNs
nonlinear fashion. Although optimality is not guaranteed, neural were investigated in the present work, considering different
networks have been applied to a variety of difficult-to-model prob- number of hidden layers (1–3), neurons (1–15), transfer functions
lems with various levels of success. (tansig or logsig), and initial random weights. Among all these net-
There are several algorithms and architectures available to use in works, the networks characterized by lowest error were selected.
an ANN model. In the present work, a feed-forward neural network The performance function is mean squared error and the training
with back propagation algorithm which is the most common net- function is Levenberg–Marquardt. For testing each network the
work structure is developed to predict the spring back. In this meth- M-fold cross validation method which is the standard technique
od the neurons are organized into layers and have only forward to obtain an estimation of a network performance over all the data,
connections. The first layer is called input layer and consists of a is used. In this method, the entire data are divided into M disjoint
number of neuron usually equal to the number of input. In general sets of equal size N/M, where N is the total data. The neural net-
each input is connected to all input neurons. The last layer is called works is trained on (M  1) subsets together and then this trained
output layer and consists of a number of computational neurons. All network is applied to the remaining subset to test the networks.
334 V. Nasrollahi, B. Arezoo / Materials and Design 36 (2012) 331–336

Table 2
The anisotropic and tensile test results.

Yield stress (MPa) UTS (MPa) Final strain Strength coeff: K (MPa) Strain hardening: n R0 R90 R45 Rm
HSLA360 381 498 0.25 881 0.227 0.59 0.87 0.90 0.82
St12 185 305 0.35 557 0.258 1.67 1.36 1.05 1.28

Table 3
Statistical comparison of FEM and experiments.

Max error Min error Average error

Percent (%) 18.2 0 4.82
Degree (Deg) 0.62 0 0.13

Fig. 6. The effect of hole type on springback in different die radii (experimental
results).

Fig. 9. The effect of hole ratio on springback for different die radii (FEM simulation
results).

Fig. 7. The effect of hole type on springback for different pad forces (experimental
results).

Fig. 10. The effect of hole ratio on springback for different materials (FEM
simulation results).

Fig. 8. The effect of hole type on springback for different materials (experimental
results).

The network is then initiated and this process is repeated M times

in a way that all the subsets are used as the testing data [24]. The
error across all M trials is computed and the network error is the
average of these M errors. In this paper, all the data are divided into
10 groups and the 10-fold cross validation is applied. Each time
20% of the data is used for testing and 10% for validation during
training to avoid overfitting. The input parameters are hole type, Fig. 11. The effect of hole ratio on springback for different number of holes (FEM
number of holes, W/L, Rd/t, Fpad/L, and rs/E. Springback is the output simulation results).
 V. Nasrollahi, B. Arezoo / Materials and Design 36 (2012) 331–336 335

Table 4
Neural Network Results for oblong holes.

Number of 1st Layer transfer 2nd Layer transfer 3rd Layer transfer Number of neuron in Number of neuron in Number of neuron in Error
hidden layer function function function 1st layer 2nd layer 3rd layer (Deg)
2 tan tan – 7 5 – 0.244
2 tan tan – 6 2 – 0.240
2 tan log – 2 5 – 0.235
2 tan log – 2 8 – 0.238
2 log tan – 2 3 – 0.229
2 log tan – 3 4 – 0.236
2 log log – 7 4 – 0.227
2 log log – 2 6 – 0.229
3 log log log 3 3 4 0.221
3 log log log 2 2 3 0.232
3 log log tan 7 4 2 0.234
3 log log tan 2 6 3 0.24
3 log tan log 2 3 2 0.223
3 log tan log 2 2 5 0.240
3 log tan tan 6 3 2 0.238
3 log tan tan 4 4 3 0.247
3 tan log log 4 5 5 0.228
3 tan log log 2 4 12 0.231
3 tan log tan 4 6 4 0.231
3 tan log tan 3 4 2 0.235
3 tan tan log 5 6 2 0.229
3 tan tan log 2 8 8 0.230
3 tan tan tan 7 3 6 0.235
3 tan tan tan 6 2 4 0.237

of the network. Because the hole type is not a digit it cannot be 5. Results and discussion
used as an input parameter. In order to solve this problem and
compare the results, two different neural network models are 5.1. Experimental results
developed as follows:
The results of bending tests are put together to evaluate the ef-
- Training an ANN just for oblong hole. fect of each parameter on springback. Fig. 6–8 show the graphs of
- Training an ANN for all type of holes by finding suitable coeffi- spring-back versus hole type in different die radii, pad forces and
cients (between 0 and 1) for each type of hole. materials respectively. These figures provide the observations sta-
ted below:
The optimum coefficient for each hole type is chosen a number
between 0 and 1 with 0.1 interval by developing the ANN with - The circle hole causes more springback than oblong and square
minimum error. holes type.

Table 5
Neural network results for all holes.

Number of 1st Layer 2nd Layer 3rd Layer Number of Number of Number of Circle Oblong Square Error
hidden layer transfer transfer transfer neuron in 1st neuron in 2nd neuron in 3rd hole hole hole (Deg)
function function function layer layer layer coeff coeff coeff
2 tan tan – 4 11 – 0.9 0.2 0.1 0.136
2 tan tan – 4 4 – 0.8 0.3 0.1 0.153
2 tan log – 4 11 – 0.9 0.4 0.2 0.144
2 tan log – 4 8 – 0.5 0.2 0.1 0.151
2 log tan – 10 5 – 0.6 0.5 0.4 0.150
2 log tan – 8 12 – 0.9 0.6 0.4 0.150
2 log log – 8 8 – 0.7 0.3 0.1 0.141
2 log log – 5 2 – 0.5 0.3 0.1 0.146
3 log log log 7 6 4 0.7 0.6 0.3 0.135
3 log log log 4 5 5 0.9 0.7 0.4 0.137
3 log log tan 10 3 6 0.9 0.7 0.5 0.141
3 log log tan 5 7 4 0.7 0.6 0.3 0.141
3 log tan log 5 8 4 0.9 0.4 0.2 0.137
3 log tan log 3 9 8 0.5 0.3 0.2 0.141
3 log tan tan 9 6 3 0.8 0.6 0.2 0.138
3 log tan tan 4 9 4 0.9 0.7 0.4 0.142
3 tan log log 5 9 6 0.5 0.4 0.3 0.138
3 tan log log 5 4 4 0.7 0.3 0.2 0.142
3 tan log tan 6 6 2 0.9 0.6 0.3 0.139
3 tan log tan 5 3 5 0.9 0.7 0.1 0.143
3 tan tan log 3 7 10 0.8 0.4 0.2 0.137
3 tan tan log 3 4 6 0.6 0.3 0.2 0.141
3 tan tan tan 3 9 9 0.6 0.3 0.1 0.138
3 tan tan tan 3 3 2 0.9 0.5 0.3 0.144
336 V. Nasrollahi, B. Arezoo / Materials and Design 36 (2012) 331–336

- The oblong hole causes more springback than the square hole. - The springback in wipe bending for parts with holes can be sim-
- The effect of hole type on springback is the same in different die ulated accurately using FEM.
radii, pad forces and materials. - The MLP, (BP) Neural networks with Levenberg–Marquardt
training function and two or three hidden layers are more
5.2. Finite element results appropriate to predict springback for perforated components.
- By employing a coefficient for each hole type the neural
The FEM results are used as training data for neural network. So networks can predict springback more accurately than using a
the first issue is to validate the FEM model by experiments. The separate network for each hole type.
FEM results are compared with over 100 experiments. The results
are summarized in Table 3. It is seen that FEM results are in good
agreement with the experiments giving an error within the accept- References
able range. This table shows that wipe bending process for parts
with holes can be simulated accurately. [1] Gardiner FJ. The springback of metals. Trans ASME 1957;79:1–9.
[2] Johnson W, Yu TX. Springback after the biaxial elastic–plastic pure bending of a
Fig. 9–11 show, some of the FEM model results. The graphs pro- rectangular plate. Int J Mech Sci 1981;23(10):619–30.
vide springback versus hole ratio in different die radii, materials [3] Johnson W, Yu TX. On the range of applicability of results for the springback of
and number of holes. These figures show that springback decreases an elastic/perfectly plastic rectangular plate after subjecting it to biaxial pure
bending – II. Int J Mech Sci 1981;23(10):631–7.
as hole ratio increases (up to hole ratio of 0.3). After that, spring- [4] Leu DK. A simplified approach for evaluating bendability and springback in
back increases are within 0.1°. It can be seen in Fig. 9 and 10 that, plastic bending of anisotropic sheet metals. J Mater Proc Technol
this is true in different die radii and materials. Fig 11 shows the 1997;66:9–17.
[5] Buranathiti T, Cao J. An effective analytical model for springback prediction in
same behavior and it indicates that by increasing the number of straight flanging processes. Int J Mater Product Technol 2004;21:137–53.
holes, springback will decrease slightly. It can also be seen that [6] Zhang D, Cui Z, Ruan X, Li Y. An analytical model for predicting springback and
more holes cause a lower springback. side wall curl of sheet after U-bending. Comput Mater Sci 2007;38(4):
707–15.
[7] Yi HK, Kim DW, Van Tyne CJ, Moon YH. Analytical prediction of springback
5.3. Neural network results based on residual differential strain during sheet metal bending. J Mech Eng Sci
2008;222:117–29.
[8] Prior AM. Applications of implicit and explicit finite element techniques to
As stated the ANN training is developed for two models. In
metal forming. J Mater Process Technol 1994;45:649–56.
Table 4 the results of ANN with least errors for different architec- [9] Finn MJ, Galbraith PC, Wu L. Use of a coupled explicit–implicit solver for
ture for oblong holes are shown. The errors are computed based calculating spring-back in automotive body panels. J Mater Process Technol
on 10-fold cross validation method. As it can be seen the error of 1995;50:395–409.
[10] Esat V, Darendeliler H, Gokler MI. Finite element analysis of springback in
this method for oblong holes are about 0.24°. In Table 5 the results bending of aluminium sheets. Mater Des 2002;23(2):223–9.
are shown for all type of holes. The coefficient for each hole type [11] Chan WM, Chew HI, Lee HP, Cheok BT. Finite element analysis of spring-back
which makes the least error is specified. It can be seen that the of V-bending sheet metal forming processes. J Mater Process Technol
2004;145:15–24.
error in this method is about 0.14° which is less than the method [12] Ling YE, Lee HP, Cheok BT. Finite element analysis of springback in L-bending
of training each hole separately. All the architectures in Table 5 of sheet metal. J Mater Process Technol 2005;168:296–302.
are accurate enough to be used. [13] Panthi SK, Ramakrishnan N, Ahmed Meraj, Singh Shambhavi S, Goel MD. Finite
element analysis of sheet metal bending process to predict the springback.
Mater Des 2010;31:657–62.
6. Conclusions [14] Inamdar MV, Date PP, Desai UB. Studies on the prediction of springback in air
vee bending of metallic sheets using an artificial neural network. J Mater
Process Technol 2000;108:45–54.
A study on modeling and predicting the springback in sheet [15] Liu W, Liu Q, Ruan F, Liang Z, Qiu H. Springback prediction for sheet metal
metals with holes on bending area is carried out, using finite forming based on GA–ANN technology. J Mater Process Technol 2007;187–
element and neural networks and compared with experimental 188:227–31.
[16] Kazan Recep, Fırat Mehmet, Tiryaki Aysun Egrisogut. Prediction of springback
results. From these results, the following conclusions can be in wipe-bending process of sheet metal using neural network. Mater Des
made: 2009;30:418–23.
[17] Farsi MA, Arezoo B. Experimental study of high-strength low-alloy sheet metal
components with holes on the bending surfaces. Proc IMechE Part B: J Eng
- The type of the hole on the bending area have an effect on the Manuf 2009;223:387–94.
springback. The circle holes have more springback compared [18] Suchy I. Handbook of die design. 2nd ed. USA: McGraw-Hill; 2006.
to oblong and the oblong holes have more springback compared [19] ASTM E 8M-04. Standard test methods for tension testing of metallic materials
[metric]. American Society for Testing and Materials; 2004.
to square holes. This is true for different die radii, pad forces and [20] ASTM E 517-00. Standard test method for plastic strain ratio R for sheet metal.
materials. American Society for Testing and Materials; 2006.
- The presence of holes causes less springback. There is a mini- [21] Haykin S. Neural networks: a comprehensive foundation. 2nd ed. Prentice Hall
International; 1998.
mum springback in 0.3 hole ratio and it is the same for different
[22] Bishop CM. Pattern recognition and machine learning. Springer; 2006.
die radii and materials. This is also true for parts with two and [23] Altinkok Necat, Koker Rasit. Neural network approach to prediction of bending
three holes. strength and hardening behaviour of particulate reinforced (Al–Si–Mg)–
- By increasing the number of holes in the part, springback will aluminium matrix composites. Mater Des 2004;25.
[24] Duda RO, Hart PE, Stork DG. Pattern classification. 2nd ed. New York
decrease slightly. (NY): Wiley Interscience; 2001.