Modeling the Controlled Rolling Critical Temperatures Using

Empirical Equations and Neural Networks

1



Antonio Augusto GORNI, Celso Gomes CAVALCANTI

Research Department, COSIPA, C.P. 11, 11573-900, Cubato SP, Brazil



Abstract

The knowledge of the critical temperatures in controlled rolling - T
nr
, Ar
3
and Ar
1
- is fundamental for the correct
selection of the temperatures where should happen the several steps of this process. In this work these parameters
were determined for the most frequently microalloyed steels processed in the Plate Mill of COSIPA. It was verified
that the isolated effect of several alloy elements over these critical parameters was not very clear, fact attributed to
their short magnitude range among the steels studied in this work. Nb and Si were the most statistically significant
elements for the determination of T
nr
and A
r3
, respectively. The mathematical modeling through neural networks for
the calculation of those two parameters, starting from steel chemical composition, was shown to be more precise than
its calculation through the former empirical models. Their performance can be further improved through the use of
larger amounts of data during the learning phase of the neural networks.

KEY WORDS: Hot Torsion Tests, No-Recrystallization Temperature, A
r3
Temperature, Mathematical Models



1
7th International Conference on Steel Rolling, The Iron and Steel Institute of J apan, Chiba, 1998, 629-633.
1. Introduction

The correct execution of the controlled rolling of a
microalloyed steel requires the knowledge of its critical
temperatures, that is, the no-recrystallization
temperatures (T
nr
or T
95%
, and T
5%
)
1)
, ferrite
transformation start (A
r3
) and finish (A
r1
). These
temperatures define the correct temperature ranges of
the several phases of controlled rolling. The roughing
stage must be conducted above T
nr
; a holding stage (no
rolling) must be adopted between T
nr
and T
5%
; the
finishing stage is delimited between T
5%
and A
r3
;
intercritical rolling occurs between A
r3
and A
r1
; and
ferrite rolling is done below A
r1
.
There are several justificatives for the determination
of that temperatures
2)
. In first place, they allow a
qualitative forecast of the evolution of the rolling loads
along temperature, that is to say, along the pass
schedule. On the other hand, its knowledge can base the
development of pass schedules that optimize the
flattening" degree of the austenite grains and/or ferrite
strain hardening.
The determination of T
nr
in laboratory was already
got by several ways: through hot torsion, compression
or even rolling tests
2)
, as well even under industrial plate
rolling conditions
3)
. The determination of Ar
3
after hot
deformation can be done using dilatometers with
coupled hot compression devices
4)
, through thermal
analysis of hot rolled samples
5)
or using a hot torsion
device with a coupled dilatometer
6)
. However, the
empirical method developed by Boratto et alii
2)
is
particularly useful, because it allows simultaneous
determination of the T
nr
, Ar
3
and Ar
1
temperatures using
only one test. Borattos approach consists of a single hot
torsion test where multiple deformation passes, applied
under decreasing temperature, are applied to the
specimen. The posterior analysis of the evolution of
mean hot strength along the inverse of the
corresponding temperature allows the determination of
T
nr
, Ar
3
and Ar
1
. However, the accuracy of these
calculated parameters depends on the subjective
interpretation of the curve, which can produce
significant variation in the final results. The same
authors studied 17 microalloyed steels using this
approach and proposed the following equation for the
calculation of T
nr
:

T C Nb Nb V V Ti Al Si
nr
= + + + + + 887 464 6445 644 732 230 890 363 357 ( ) ( )

(1)

An option for the calculation of T
nr
, with a more
sound theoretical foundation, requires the use of
equations that describe recrystallization and
precipitation kinetics. In this approach, T
nr
is defined as
the intersection of the curves corresponding to the time
necessary for the occurrence of 95% of austenite
recrystallization and for the occurrence of 5% of
carbonitride precipitation, as shown in Figure 1. Such
model was originally developed by Dutta and Sellars
1)

and refined by other authors
7)
. The time required for 5%
of precipitation can be determined by the following
equation:

t
R T
Nb Z R T T k
p
R
s
0 05
10
3 2
200
200000
270000 25 10
,
exp
exp exp
, .
(ln )
=

(2)
700
800
900
1000
1100
0,1 1 10 100 1000 10000
Time [s]
T
e
m
p
e
r
a
t
u
r
e

[

C
]
5% Rex
(t
x0.05
)
95% Rex
(t
x0.95
)
5% Ppt
(t
p0.05
)
T
nr
ou T
95%
T
5%

Figure 1: Interaction between austenite recrystallization
and precipitation kinetic curves, showing as
the different no-recrystallization temperatures
(T
nr
/T
95%
and T
5%
) are determined.

where T
R
is the temperature under which the material
was deformed, is the true strain, Z is the Zener-
Hollomon parameter, R is the universal gas constant, T
the temperature and k
s
the rate of Nb supersaturation,
given by

k
Nb C
N
s
T
=
+

[ ]
,
12
14
10
2 26
6770

(3)

and the time necessary for 5% recrystallization can be
calculated by the formula

t d
R T T
Nb
x0 05
20
0
2
675 10
300000 275000
185
,
, . exp exp =

(4)

where d
0
is the initial austenitic grain size of the
specimen.
This method has a particular advantage: it not only
allows the calculation of T
nr
, that is, the minimum
temperature at which austenite completely recrystallizes
between hot rolling passes, as well it permits the
calculation of T
5%
, the maximum temperature where
austenite did not show any recrystallization after hot
deformation, which corresponds to the intersection of
t
p0.05
and t
x0.95
. This last parameter can be calculated by
the formula

05 . 0 95 . 0
66 . 7
x x
t t =
(5)

The Ar
3
temperature values got by Boratto et al.
2)

showed good fitting with the formula proposed by
Ouchi et al.
5)
:

) 8 , 0 ( 35 , 0 80 15 20 80 310 910
3
+ = h Mo Cr Cu Mn C Ar
(6)

where h is the thickness of the sample being submitted
to rolling. Actually, the last portion of this equation is a
form of correcting variations in the cooling rate of the
test sample. This equation was developed using thermal
analysis data collected during the air cooling of hot
rolled samples
5)
.
Up to this moment there is very little information
about the quantitative influence of alloy elements over
the Ar
1
temperature, so no model was proposed for its
calculation
2)
.
The aim of this work was to determine the T
nr
, Ar
3

and Ar
1
temperatures for the most processed
microalloyed steels at the COSIPA plate mill, as well to
develop mathematical models for its calculation from
the plate chemical composition.


2. Experimental Procedure

Hot torsion tests, using the methodology proposed
by Boratto et al.
2)
, were carried out at the Materials
Engineering Department of the Federal University of
So Carlos, in So Carlos, Brazil. Ten microalloyed
steels were studied, whose chemical compositions are
shown in Table I.

C Mn Si Al Cr Cu Nb V Ti N
N1 0,18 1,34 0,30 0,025 - - 0,033 - - 0,0074
N2 0,14 1,02 0,27 0,035 - - 0,020 - - N/A.
N3 0,15 0,77 - 0,039 - - 0,014 - - N/A.
NT1 0,14 1,11 0,30 0,044 - - 0,020 - 0,015 0,0054
NT2 0,14 1,34 0,23 0,035 - - 0,033 - 0,014 0,0048
NT3 0,10 1,12 0,30 0,040 - - 0,013 0,020 - N/A.
NT4 0,10 1,16 0,33 0,027 - - 0,035 0,023 - N/A.
NTV 0,12 1,50 0,31 0,038 - - 0,047 0,051 0,020 0,0064
NCC1 0,16 1,03 0,41 0,029 0,54 0,23 0,025 - - 0,0107
NCC2 0,13 0,99 0,38 0,042 0,50 0,22 0,014 - - 0,0095

Table I: Chemical composition of the steels studied in
this work, weight percent.

The hot torsion test specimens were firstly heated up to
1150
o
C under a heating rate of approximately 1.7
o
C/s.
They were kept at this temperature for ten minutes.
After this soaking stage, the specimens were cooled
under a rate of 0.5
o
C/s down to 1050C. From this point
on, the cooling rate was increased to 1
o
C/s.
Simultaneously, the hot torsion test effectively began: a
hot deformation pass was applied every 20 seconds,
with a strain degree of 0.2 and strain rate of 1.0 s
-1
.
Therefore, the passes had an interval of 20C between
each other. It was aimed to apply a total of 20
deformation passes, so the last pass was applied under a
temperature of approximately 670C. This procedure
was repeated five times for each steel studied. The
results of these tests allowed the drawing of graphs
between mean hot strength versus the inverse of the
corresponding hot deformation temperature. The
analysis of these graphics, according to the
methodology defined by Boratto et al.
2)
, allowed the
determination of T
nr
, Ar
3
and Ar
1
temperature values.
The statistical analysis for the determination of the
correlation between these critical temperatures and the
corresponding chemical compositions of the studied
steels was carried out using the software Systat.
A comparison of the values of T
nr
calculated by
equation (1) and (2) to (4) was also done, as well the
calculation of Ar
3
values by equation (6).
Finally, the neural network technique was used for
the development of models for the calculation of T
nr
and
Ar
3
from the chemical composition of steels. Given the
small amount of data available for these temperatures
only ten cases this approach can not be used for the
direct determination of these critical temperatures, as it
requires a much greater amount of data for training and
testing of the models. Instead of that, neural network
models were developed for the determination of the
mean hot strength versus the inverse of the hot
deformation temperature graphs, starting from the
chemical composition of the steel and assuming the
same hot forming conditions described above. From that
graph, T
nr
and Ar
3
can be determined using the approach
of Boratto et al.
2)
.
A previous performance analysis showed that best
results were got using two individual neural networks.
The first network generates a mean hot strength versus
inverse of hot deformation temperature graph limited to
the 1050 to 780
o
C temperature range; the second was
developed for use in the 900 to 720
o
C range. The graphs
they generated are respectively used for the
determination of T
nr
and Ar
3
.
The neural networks used in this work were of the
Rummelhart type, with three neuron layers. The first
layer was used for the input of the necessary data for the
calculations: the amounts of C, Mn, Si, Nb, Ti, V, Cr,
Cu, Al and N, as well the deformation temperature. The
second layer, also known as the hidden layer, is used to
improve the learning capacity of the neural network. It
had 23 neurons, according to the Hecht-Kolmogorov
theorem
8)
, which proposes that the number of neurons in
the hidden layer must be to the double of the neurons in
the input layer (in this case, 11) plus one. The output
layer had just one neuron, which is equal to the mean
hot strength corresponding to the input data. The
training and testing of these neural networks was done
using the NeuralWorks software.
The evaluation of the statistical and neural networks
models developed in this work was done comparing the
corresponding values of the Pearson correlation
coefficient r and the standard error of estimate.


3. RESULTS AND DISCUSSION

The results of T
nr
, Ar
3
and Ar
1
got for the steels
studied in this work are shown in table II. The
determination of Ar
1
was not possible for all samples,
due to the premature breaking of a significant fraction
of the hot torsion specimens before reaching the range
of temperatures corresponding to that parameter. The
resulting lack of data prevented the determination of the
correlation between Ar
1
and the chemical composition
of the steels.

T
nr
Ar
3
Ar
1

N1 868 730 -
N2 894 775 -
N3 879 712 -
NT1 861 772 -
NT2 899 753 -
NT3 840 760 -
NT4 918 748 -
NTV 916 754 697
NCC1 883 748 708
NCC2 894 754 707

Table II: Controlled rolling critical temperatures
determined for the studied steels.

The results of the Pearson correlation coefficient
showed that only Nb had significant effect on T
nr
, with r
equal to 0,634. The results of the principal component
analysis showed that Mn, Nb and V are directly
correlated to T
nr
, as showed in Figure 2.

-1
-0,5
0
0,5
1
-1 -0,5 0 0,5 1 1,5
Factor 1
F
a
c
t
o
r

2
Nb
Mn
Cr
Cu
C
eq
Ti Si
V
T
nr
Al
Ar
3
C

Figure 2: Principal component analysis graph between
T
nr
and the chemical composition of the
studied steels.

The effect of Nb over T
nr
was thoroughly checked in
the literature
2)
; for its turn, the influence of Mn and V
detected here apparently stems from a statistical
correlation between the amounts of those elements and
Nb. In fact, the Pearson correlation coefficient r
between the amounts of Mn and Nb is equal to 0.845;
this same parameter equals 0.672 between the amounts
of Mn and V. The principal component analysis also
showed a light negative influence of Al on T
nr
.
Apparently only Si had a significant influence over
Ar
3
; the value of r corresponding to that correlation was
about 0.600. Even the principal component analysis did
not present any significant correlation. Possible reasons
for this behavior can be associated to the few number of
analyzed steels, as well the corresponding narrow range
of alloy elements amounts.
It is interesting to notice the peculiar results that
were obtained for the N3 steel, which had the minimum
amount of microalloying elements. Surprisingly, its T
nr

is not between the smallest ones. A possible explanation
for that fact is the absence of Si in this steel. This
element, according to the equation of Boratto
2)
, tends to
decrease T
nr
. Its absence, therefore, can exert an inverse
effect. The same fact can have some relationship with
the low value of Ar
3
determined for that alloy, as the
presence of Si tends to increase that parameter
9)
.
Data in Table III allows a comparison between T
nr

values measured in this work and calculated by the
models of Boratto, Dutta and by the neural network
developed here. Only steels with known N amount were
considered for this analysis. This data reveals that the
neural network had the best forecasting performance,
while Dutta and Boratto models showed a slight greater
error.


Steel
T
nr

Boratto
[C]
T
nr

Boratto
[C]
T
nr
or T
95%

Dutta
[C]
T
nr

Dutta
[C]
T
nr

NN
[C]
T
nr
NN
[C]
Nb1 886 18 881 13 872 4
NbTi1 887 26 881 20 877 16
NbTi2 898 -1 902 3 897 -2
NbTiV 895 -21 905 -11 911 -5
NbCrCu1 878 -5 880 -3 892 9
NbCrCu2 878 -16 872 -22 873 -21
S.E.E. 20 17 14

Table III: Values of T
nr
s calculated using the several
models described in this paper: Boratto, Dutta
and the neural network.

Firstly the performance of the Boratto equation will
be analyzed. This formula was originally developed
using data from almost fully solubilized microalloyed
steels, which were austenitized at 1260C before the hot
torsion tests. Therefore, T
nr
values calculated by this
model in table III considered the corresponding
solubilized amount of microalloying elements at the
austenitization temperature used in this work, that is,
1150C, calculated by a thermodynamical solubilization
model for multimicroalloyed steels
10)
. Such correction
significantly increased the precision of the T
nr
values
calculated by the Borattos model.
One can observe that Borattos model presented
tendency to overestimate the values of T
nr
, except for
the steels NbTiV and NbCrCu. In the last alloy, this fact
can be attributed to the presence of Cu, once there is
some evidence that this element also restricts austenite
recrystallization, contributing to the elevation of T
nr
11)
.
Other deviations observed between Borattos
equation and the experimental data got in this work can
be attributed to experimental dispersion and the
different methodologies adopted in the tests of both
works.
The model of Dutta also tended to overestimate the
values of T
nr
(that is to say, T
95%
), but this was not true
for the NbTiV and NbCrCu steels. That seems to be an
additional indication of the effect of Cu over T
nr

mentioned before.
The neural network model learns the relationships
between data by itself, dispensing the previous
definition of a relationship between the several
variables. The values of T
nr
, calculated from the graphs
average hot strength versus inverse of temperature,
determined by the neural network, were slightly higher
than the measured values. It is curious to verify that, in
spite of that model based its calculations directly from
the experimental results, it also significantly
underestimates the T
nr
value in one of the NbCrCu
steels. This case also corresponded to the maximum
deviation between measured and neural network
calculated T
nr
values: 21C.
On the other hand, it must be noted that the neural
network calculated T
nr
values were determined by an
indirect way. That is, this model did not supply directly
the calculated T
nr
values, but instead the graph between
mean hot strength versus the inverse of temperature,
which is used to determine T
nr
using Borattos
methodology. Once that determination frequently
involves subjective judgement for the analysis of the
curves, particularly when the intersection between the
two straight lines is ill conditioned, its precision can be
harmed by random errors. Unfortunately, as T
nr
data is
scarce in this work, the development of a neural
network model for its direct determination from steel
chemical composition is not possible at this moment.
However, this still is a possibility in the future, if more
data becomes available.
Table IV shows the result of the comparison
between experimental and calculated values of Ar
3
. Two
models were used for the calculation of this parameter:
Ouchi and the neural network.


Steel
A
r3

Ouchi
[C]
A
r3

Ouchi
[C]
A
r3

NN
[C]
A
r3

NN
[C]
Nb1 740 10 728 -2
NbTi1 762 -10 761 -10
NbTi2 748 -5 756 3
NbTiV 744 -10 758 4
NbCrCu1 753 5 754 6
NbCrCu2 764 10 753 -1
E.P.E. 10 7

Table IV: Values of Ar
3
calculated using the Ouchi and
neural network models.

In the same way as observed for T
nr
, only steels with
known amount of N were considered in this analysis.
The neural network model showed slight better
performance than Ouchis formula. However, in both
cases, Ar
3
determination was more precise than T
nr
.
Besides that, both Ar
3
models showed random errors,
with no tendency to over or underestimate the
experimental results.
Data scarcity problems that affected the T
nr
neural
network model also occurred during the development of
the Ar
3
neural network model. This problem will be
solved as more data becomes available. In fact, some
papers about the development of neural networks
models for the calculation of Ar
3
were already
published, although they considered data from
conventional CCT diagrams, where steel is not
previously hot formed before cooling
12,13)
.


4. CONCLUSIONS

This work about mathematical models for the
calculation of the critical temperatures of controlled
rolling (T
nr
, Ar
3
and Ar
1
) from steel chemical
composition arrived to the following conclusions:

The effects of the chemical composition over Ar
1

could not be investigated due to the premature
breaking of the hot torsion test specimens submitted
to multiple passes;

The element with largest influence over T
nr
was Nb;
Mn and V presented considerable influence, but
there are doubts if their effects are legitimate or if
they derive from a correlation with the Nb amount
of the steels studied in this work. Apparently Si
indeed tends to decrease T
nr
, as observed in one of
the alloys. It was also verified a slight negative
influence of Al;

A statistical analysis about alloy elements influence
over Ar
3
showed that only Si exerted a significant
influence. This fact was confirmed by the low Ar
3

value corresponding to a particular alloy which did
not included this element;

The comparison between several mathematical
models used for the calculation of T
nr
showed that
their performance is better along the following
sequence: Boratto Dutta Neural Networks;

Also the neural network model for calculation of
Ar
3
presented better performance than an empirical
model, in this case, the Ouchi formula;

A serious limitation of the neural network models
developed in this work is the narrow range of alloy
element amounts present in the data sets used for its
training. This fact severely limits the chemical
composition input values that can be used for the
calculation. A more general and precise neural
network model for T
nr
and Ar
3
calculation would
demand data of dozens of alloys.


5. ACKNOWLEDGEMENTS

The authors would like to express their gratefulness to
Marcos Antonio Stuart and Jos Herbert Dolabella
da Silveira, from the Plate Mill Department at
Companhia Siderrgica Paulista - COSIPA, for their
vital support in the conception and preliminary activities
of this work. The contribution of Prof. Dr. Oscar
Balancin and its staff, at the Department of Materials
Engineering in the Federal University of So Carlos
(UFSCar), is gratefully acknowledged, especially during
the execution of the hot torsion tests.


Nomenclature

T: Temperature (K)
R: Universal Gas Constant (J /K mol)
Z: Zener-Hollomon Parameter
: True Strain
k
s
: Niobium Supersaturation Ratio
d
o
: Initial Austenitic Grain Size (m)
t: Time (s)
h: Thickness (mm)


References

1) DUTTA, B. & SELLARS, C.M. Materials Science and Technology,
March 1987, 197-206.

2) BORATTO, F. et al. In: THERMEC 88. Proceedings. Iron and
Steel Institute of J apan, Tokyo, 1988, p. 383-390.

3) MACCAGNO, T.M. et al. ISIJ International, November 1994, 917-
922.

4) LOTTER, V. et al. Thyssen Technische Berichte, 1984, 13-26.

5) OUCHI, C. et al. Transactions of the ISIJ, 22(3):214-222, March
1982.

6) ROBILLER, G. et al. Stahl und Eisen, 9 Februar 1987, 111-114.

7) BAI, D.Q. et al. Metallurgical Transactions A, October 1993, 2151-
2159.

8) HECHT-NIELSEN, R. Neurocomputing. Addison-Wesley
Publishing Company, Reading, 1991. 433 p.

9) CHOQUET, P. et al. In: INTERNATIONAL CONFERENCE ON
STRENGTH OF METALS AND ALLOYS. Proceedings.
Canadian Institute of Metallurgy, 1985.

10) GORNI, A.A. In: Spreadsheets in Science and Engineering.
Springer-Verlag, Heidelberg, 1998. 289 p.

11) ABE, T. et al. Transactions of the ISIJ, 1987, 478-84.

12) DONADILLE, C. e al. La Revue de Metallurgie CIT, Oct. 1992,
892-894

13) VERMEULEN, W. et al. Steel Research, Feb. 1997, 72-79.