2011 Seventh International Conference on Natural Computation

Study on a Zerilli-Armstrong and an artificial

neural network model for 4Cr5MoSiV1
Quenched Steel at High Strain Rate
Jing Wang
Changzhou Institute of Light Industry Technology
Changzhou, China

with adaptive population size. And the artificial neural

Abstract: The flow stress in compression of 4Cr5MoSiV1
network (ANN) is used in establishing of constitutive
quenched Steel was investigated by means of a split Hopkinson
predicted model. The accuracy of these two constitutive
pressure bar (SHPB) experiment apparatus under different
models is higher than the traditional regression algorithm. So
temperature. According to the stress-strain curves, the
these two models can be used to finite element simulation in
Zerilli-Armstrong constitutive model was chosen and its
the process of quenched steel high speed cutting, and can
relationship parameters were determined by Genetic
improve the simulation precision.
Algorithm (GA) with adaptive population size. At the same
time, the Back Propagation artificial neural network (BP II. EXPERIMENT MATERIAL AND METHOD
ANN) was used for establishing constitutive model of
The hot work mold steel 4Cr5MoSiV1 is the experiment
4Cr5MoSiV1 quenched Steel. Compared with the
material. The heat treatment of vacuum quenching was used
experimental data, two constitutive equations can predict the
and the hardness attained HRC47-49.
flow stress very well, and the prediction method using the BP
artificial neural network had higher accuracy. The research The dynamics compression test was carried out in the
provides a necessary material characteristic parameters for SHPB experiment apparatus and produced the data about
finite element numerical simulation of 4Cr5MoSiV1 quenched stress, strain, strain rate and temperature of the material. It
Steel, and the two prediction method can be widely used to was finished in two compression bars whose diameters were
establish other nonlinear relations of manufacturing 12.7mm and 5mm. The lengths of projectiles were 490mm
procedure. and 190mm. The test samples were two kinds of cylinder,
one diameter and length is 5mm, the other is 1mm. the true
stress and true strain can be produced by the test under
Key words：4Cr5MoSiV1 Quenched steel, SHPB, high strain
different temperature ranging from 20℃ to 600℃ with a
rate, Zerilli-Armstrong model ， artificial neural network,
strain rate of 1000S-1 to 10000 S-1.
Genetic Algorithm
The stress-strain curve shown in Figure 1 was attained from
I. INTRODUCTION
the SHPB compression test which carried out under different
As a typical hot work mold steel with good property, temperature and strain rate. From this figure, the flow stress
4Cr5MoSiV1 is used in casting die, forging die and of 4Cr5MoSiV1quenched steel increased under lower
extrusion die and so on. The steel has well comprehensive temperature and higher strain rate, and decreased under
properties, especially the high temperature strength and higher temperature and lower strain rate. This shows that the
thermal fatigue performance. In this present work, the deformation of 4Cr5MoSiV1 is with complex dynamic
dynamics property of 4Cr5MoSiV1 quenched steel was response characteristics under high strain and wide
investigated by means of a split Hopkinson pressure bar temperature range. It is the result of existing and competition
(SHPB) experiment apparatus. According to the data of of hardening dynamic recovery and dynamic
stress, strain, strain rate and temperature produced by recrystallization mechanism.
experiment, the Zerilli-Armstorng (Z-A) experience equation
is chosen as the constitutive model. Its material parameters
of Z-A model are optimized by Genetic Algorithm (GA)

978-1-4244-9953-3/11/$26.00 ©2011 IEEE 247

 The Z-A constitutive equation achieved based on
generics algorithm is shown in Equation (2).

σ = 500 + 2000 exp( − 0.0125T + 0.00136T ln ε ) + 500ε 1.1923 (2)

Compared the value achieved from the Equation (2) with
the experiment data, get the results shown in Figure2.From
the Figure2, the predicted value is good agreement with
experimental data. Under 7000S-1- 473K, the maximum
positive error between predicted and experimental is +5.84%,
the maximum negative error between predicted and
experimental is 0.
Figure 1. Stress-strain curve under the conditions of different
d diff i

III. PREDICTION OF ZERILLI-ARMSTRONG CONSTITUTIVE

MODEL

A. Zerilli-armstrong constitutive equation

The material dynamics constitutive relationship is that the
flow stress is respond to thermo-dynamical state which is
composed by thermo-dynamical parameters such as strain,
strain rate and temperature. In studying of mechanical
properties of material under high strain rate, the effect of
strain rate and temperature isn’t ignored. So, the constitutive Figure 2. Flow stress using Z-A model predict and experimental data

relationship is defined as heat viscoplastic constitutive

relationship. Under 6000S-1-673K, the maximum positive error between
predicted and experimental is +1.165%, the maximum
At present, the main heat viscoplastic constitutive models negative error between predicted and experimental is
are: Zerilli-armstrong model, Johnson-cook model, -3.18%; Under 5000S-1-873K, the maximum positive error
Follansbee-kocks model, Bodner-paton model and so on. By between predicted and experimental is +1.81%, the
contrast, the Zerrilli-Armstrong model has the different maximum negative error between predicted and experimental
expression to different crystal structure. It is suited for is -5.74%。
body-centered cube model and face-centered cube metal. The
crystal structure of 4Cr5MoSiV1 is body-centered cube, so IV. BACK PROPAGATION NEURAL NETWORK MODELING
its expression is shown In order to describe the relationship between stress and
strain, strain rate, temperature, the experience equation must
σ e = C0 + C1 ⋅ exp [ −C3T + C 4T ln ε e ] + C5ε e n
(1) be chosen and various material parameters must be
determined. The neural network can conquer the
Where σ e is flow stress ； ε e is equivalent strain ； ε e is disadvantage of these methods and don’t need the equation

equivalent strain rate; T is temperature; C0, C1, C3, C4, and the material parameters. After training the network using

C5and n are material parameters. the experimental data, the constitutive model can be saved in
the structure of neural network. In this present, BP neural
B. Equation factor solved by genetic algorithm network is used to develop the constitutive model of
The genetic algorithm with adaptive population size was 4Cr5MoSiV1 quenched steel.
used to get the six material parameters of Z-A model.
A. Development of constitutive model based on back
Through the GAOT toolbox of MATLAB, the six material
propagation artificial neural network.
parameters can be optimized.
Considering the three strain rates and three temperatures,
C. Result and validation

248
the equivalent plastic strain ε , temperature T, and strain The compared curve between the experimental value and
rate ε are selected as the input of BP neural network, the the predicted value based on the neural network is shown in
stress σ as the output, ten neurons as the hidden layer. Figure 3. From this curve we can see the precision of neural
network is high and the error between the calculated value
B. Function expression of neurons.
and experimental value about flow stress is small. Under the
The structure of output layer neurons is exactly the same
condition of 7000S-1-473K, the maximum positive error
with that of the hidden layer neurons, but the transfer
between predicted and experimental is +0.858%, the
function is different. The output function expression of the
maximum negative error between predicted and experimental
jth neurons of hidden layer is described
is -0.53%; Under the condition of 6000S-1-673K,the

z = f (∑
n
w * xi + b )
maximum positive error between predicted and experimental
j j ,i j
j = 1,2, ,10 (3)
i=1
is +0.829%, the maximum negative error between predicted
Where j is the jth neurons of hidden layer; and experimental is -0.952% ； Under the condition of
W j = (w j ,1 w j , 2 … w j , n ) 5000S-1 -873K, the maximum positive error between
is the weight factor of the jth neuron to
predicted and experimental is +1.139%, the maximum
n inputs; bj is threshold of the jth neuron. The weight and the negative error between predicted and experimental is
threshold need to be optimized. is the ith input x i (i = 1, , n) -1.131%。
to hidden layer neurons. In this network, n=3, that is
( x1 x 2 x 3 ) = (ε T ε )
and zj is the stress σ . f is the transfer
function of neurons. Here, the tansig() function is chosen as
the transfer function of hidden layer neurons purlin()
function is chosen for output neurons.

C. Function expression of BP neural network about

4Cr5MoSiV1 constitution equation.
According to tansig(), purlin(), and n=3, the output
expression of BP neural network is described as

10 ⎛ 1 − e −2×[ w j ,1×ε + w j , 2×T +w j , 3×ε +b j ] ⎞

σ = ∑ lw j * ⎜⎜ ⎟ + bs
−2×[ w j ,1×ε + w j , 2 ×T + w j , 3×ε +b j ] ⎟
(4) Figure3. Comparison of flow stress using BP model predict with
j =1 ⎝1 + e ⎠ experimental data

LW = [ lw1 , lw2 , lw3 , lw4 , lw5 , lw6 , lw7 , lw8 , lw9 , lw10 ]
Where V. CONCLUSION
express the connection weight from hidden layer to output (1) Through the experiment of SHPB, the data of stress,
layer in BP neural network. bs is the threshold of output strain, strain rate and temperature about 4Cr5MoSiV1 can be
neurons. attained. The experiment shows that the lower temperature
D. Optimization of parameters in neural network and the higher strain rate will increase the flow stress of
equation 4Cr5MoSiV1.

In this present, the experimental data are chosen as the (2) The adaptive genetic algorithm is used to optimize the
training data of BP neural network. After the structure parameters of Z-A model and get the constitutive equation
designing of BP neural network, the function trainlm of with high precision.
Leverberg-Marquart method is chosen to train the network in
(3) BP neural network is used to establish the dynamics
software of MATLAB. We need to define the maximum
constitutive equation with high strain rate of 4Cr5MoSiV1.
number of iterations and minimum mean-squared error as the
The structure of neural network is 3-10-1. The input
terminated condition. This trained BP neural network is used
parameters are equivalent plastic strain ε , temperature T,
to get the fitting data.
strain rate ε ; the output parameter is true stress σ ; the hidden
E. Result and validation layer is composited with 10 neurons.

249
 (4) The precisions of the two prediction constitutive experiments and modeling, Mechanics of Materials, 2001, 33(8) :

models are high. The one of the model using BP neural 425-439

network is higher. So this method can be promoted to other [4] Qiang Li, Y. B. Xu and M. N. Bassim. Dynamic mechanical behavior

nonlinear process modeling. of pure titanium. Journal of Materials Processing Technology, 2004,
155-156 (30):1889-1892
REFERENCES
[5] XU Wen-chen,SHAN De-bin,Lü Yan. Prediction of flow stress of
[1] CHENG Qun-lin, KE Ying-lin, DONG Hui-yue etac. Simulation of BT20 titanium alloy by BP neural network. Ordnance Material Science
chip formation in high-speed hard machining. Journal of Zhejiang and Engineering, 2007,30(3):33-36
University(Engineering science),2007,41 (3): 509-513 [6] S.H. Mousavi Anijdan, H.R. Madaah-Hosseini and A. Bahrami. Flow
[2] Songwon Seo, Oakkey Min and Hyunmo Yang et al. Constitutive stress optimization for 304 stainless steel under cold and warm
equation for Ti–6Al–4V at high temperatures measured using the SHPB compression by artificial neural network and genetic algorithm. Materials
technique. International Journal of Impact Engineering, 2005, & Design, 2007, 28(2): 609-615
31(6):735-754 [7] YANG Yong; KE Ying-lin; DONG Hui-yue. Constitutive model of
[3] Sia Nemat-Nasser, Wei-Guo Guo, Vitali F. Nesterenko et al. Dynamic aviation aluminum-alloy material in metal machining. The Chinese
response of conventional and hot isostatically pressed Ti–6Al–4V alloys: Journal of Nonferrous Metals, 2005,15(6):854-859

250