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

net/publication/228365925

A solidification heat transfer model and a neural network based algorithm

applied to the continuous casting of steel billets and blooms

Article  in  Modelling and Simulation in Materials Science and Engineering · October 2005

DOI: 10.1088/0965-0393/13/7/005

CITATIONS READS

23 157

5 authors, including:

C. A. M. dos Santos Eugenio Liborio Feitosa Fortaleza

University of São Paulo University of Brasília
109 PUBLICATIONS   898 CITATIONS    42 PUBLICATIONS   85 CITATIONS   

SEE PROFILE SEE PROFILE

Carlos Raimundo Frick Ferreira A. Garcia

Universidade Federal do Rio Grande do Sul University of Campinas
13 PUBLICATIONS   68 CITATIONS    331 PUBLICATIONS   6,070 CITATIONS   

SEE PROFILE SEE PROFILE

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

Correlations between study Thermal Variables Solidification, Microstructure and Resistance to Fatigue in Al-Ni alloys. View project

Mg alloys View project

All content following this page was uploaded by Eugenio Liborio Feitosa Fortaleza on 18 February 2015.

The user has requested enhancement of the downloaded file.

INSTITUTE OF PHYSICS PUBLISHING MODELLING AND SIMULATION IN MATERIALS SCIENCE AND ENGINEERING
Modelling Simul. Mater. Sci. Eng. 13 (2005) 1071–1087 doi:10.1088/0965-0393/13/7/005

A solidification heat transfer model and a neural

network based algorithm applied to the continuous
casting of steel billets and blooms
Carlos A Santos1 , Eugênio L Fortaleza1 , Carlos R F Ferreira2 ,
Jaime A Spim2 and Amauri Garcia1
1 Department of Materials Engineering, State University of Campinas, UNICAMP, 13083-970,

Campinas, SP, Brazil

2 Foundry Laboratory, Federal University of Rio Grande do Sul, Brazil

E-mail: [email protected]

Received 12 January 2005, in final form 9 August 2005

Published 12 September 2005
Online at stacks.iop.org/MSMSE/13/1071

Abstract
This work presents the development of a computational algorithm applied to
improve the thermal behaviour in the secondary cooling zone of steel billets
and blooms produced by continuous casting. A mathematical solidification
heat transfer model works integrated with a neural network based algorithm
(NNBA) connected to a knowledge base of boundary conditions of operational
parameters and metallurgical constraints. The improved strategy selects a
set of cooling conditions (in the secondary cooling zone) and metallurgical
criteria established to attain high product quality, which are related to a more
homogeneous thermal behaviour during solidification. Initially, the results of
simulations performed by using the mathematical model are validated against
experimental industrial data, and good agreement is observed, in any case
examined, permitting the determination of nominal heat transfer conditions
by the inverse heat conduction method. By using the numerical model linked
to a NNBA results have been produced determining a set of casting conditions,
which has permitted better strand surface temperature profile and metallurgical
length to be attained during the continuous casting of SAE 1007 billets and
SAE 1025 blooms.

1. Introduction

The continuous casting of steel is one among a number of industrial processes that is subjected
to perturbations during continuous operation. In this case, it is convenient to develop a control
system connected to a heat transfer mathematical model, which must be able to analyse the
solidification progress. Modelling has the ability to simulate operating parameters, which

0965-0393/05/071071+17$30.00 © 2005 IOP Publishing Ltd Printed in the UK 1071

1072 C A Santos et al

Figure 1. Schematic representation of the continuous casting process.

must be continuously compared with a system of metallurgical rules in order to guarantee that
optimal conditions for manufacture are attained. The purpose of continuous casting simulation
is to provide a processing map that is generally a temperature against the time plot of the process
or distance from the meniscus [1–6]. Figure 1 shows a schematic representation of a continuous
caster and the different cooling zones along the machine. The casters have been implemented
with modern equipment for billets, slabs or blooms, multiple casting and process control.
The idea of using simulation to improve continuous casting processes is not just a
theoretical concept, and its practicality has already been demonstrated [7–15]. This paper
describes a software, which has been developed based on the interaction between a finite
difference heat transfer solidification model, a neural network based algorithm (NNBA) and
a knowledge base of metallurgical constraints which is explored by using heuristic search
techniques. The heat transfer model is validated against industrial continuous casting results
concerning the production of SAE 1007 steel billets and SAE 1025 steel blooms. The
software has been used to explore the space parameters settings in order to find the best
cooling conditions, which result in lower reheating between spray cooling zones, complete
solidification just before the unbending point and temperatures at the unbending and flame
cut-off points inside a recommended range.

2. Mathematical solidification heat transfer model

The mathematical formulation of heat transfer to predict the temperature distribution during
solidification is based on the general equation of heat conduction in the unsteady state, which
is given for two-dimensional heat flux by [16]



∂ ∂ ∂T ∂ ∂T o
(ρ · c · T ) = kx + ky + q, (1)
∂t ∂x ∂x ∂y ∂y
where ρ is density [kg m−3 ]; c is specific heat [J kg−1 K−1 ]; k is thermal conductivity [W m−1
K−1 ]; ∂T /∂t is cooling
o
rate [K s−1 ], T is temperature [K], t is time [s], x and y are space
q
coordinates [m] and represents the term associated to the internal heat generation due to the
phase change. It was assumed that the thermal conductivity, density and specific heat vary
with temperature. The co-ordinate system is one moving at the casting speed with x and y
orthogonal to the casting direction.
Continuous casting of steel billets and blooms 1073

Approximating equation (1) by finite-difference terms [17], we obtain:

 n 
n+1
Ti,j − Ti,j
n
Ti+1,j − 2Ti,j
n n
+ Ti−1,j n
Ti,j +1 − 2Ti,j + Ti,j −1
n n
o
ρ·c· =k· + + q, (2)
T x 2 y 2
where i, j are the element positions according to ‘x’ and ‘y’ axes, n and n + 1 refer to
temperatures before and after the incremental time interval t, respectively, and the stability
criteria are given by t < [(x 2 /2α)+(y 2 /2α)] where α = k/ρ · c is the thermal diffusivity
[m2 s−1 ].

2.1. Phase change

In this study, a fixed grid methodology is used with a heat source term due to phase
transformation, which is given by an explicit solid fraction–temperature relationship. The
solid fraction depends on a number of parameters; however, it is quite reasonable to assume fs
varying only with temperature, where fs is the solid fraction and L is the latent heat of fusion
[J kg−1 ]. So, we can write c
= c − L · ∂fs /∂T , and the term (L · ∂fs /∂T ) can be considered
a pseudo-specific heat. At the range of temperatures where solidification occurs for metallic
alloys, the physical properties will be evaluated taking into account the amount of liquid and
solid that coexists in equilibrium at each temperature, as shown by equations (6)–(8) [18]. The
sub-indices S and L indicate solid and liquid, respectively. For carbon steels, fs is appropriately
described by the lever rule, given by
1 TL − T
fS = · , (3)
1 − k o TF − T
where ko is the partition coefficient, TL is the liquidus temperature and TF is the solvent fusion
temperature.

2.2. Analogy between thermal systems and electrical circuits

Multiplying equation (2) by x · y · z, where At = y · z, we obtain [19]
 n 
n+1
Ti,j − Ti,j
n
(Ti+1,j − 2Ti,j
n n
+ Ti−1,j n
) (Ti,j −1 − 2Ti,j + Ti,j +1 )
n n
At · x · ρ · c
· = At · k· + ,
t x y
(4)
where At is the area of element i, j [m2 ]. By using an analogy between a thermal system and
the passive elements of an electrical circuit, the thermal capacitance (CT ) represents the energy
accumulated in a volume element i, j from the grid and is given by CTi,j = x · y · z · ρ · c
,
where x · y · z is the volume of the element i, j [m3 ].
The thermal flux between central points has a thermal resistance at the heat flux line
(RT ) from point i + 1, j or i − 1, j to point i, j or i, j + 1 or i, j − 1 to i, j given by
RT = (x/k · At ) or (y/k · At ), where x and y correspond to the distance between
central points of elements [m].
Then, equation (4) becomes
 n n n n 
Ti+1,j Ti−1,j Ti,j +1 Ti,j −1
n+1
Ti,j = t · + + +
τ(i+1,j ),(i,j ) τ(i−1,j ),(i,j ) τ(i,j +1),(i,j ) τ(i,j −1),(i,j )


t
+ 1− · Ti,j
n
, (5)
τ(i,j ),(i,j )
1074 C A Santos et al

where
τ(i+1,j ),(i,j ) = cTi,j (RTi+1,j + RTi,j ), τ(i−1,j ),(i,j ) = cTi,j (RTi−1,j + RTi,j ),
τ(i,j +1),(i,j ) = cTi,j (RTi,j +1 + RTi,j ), τ(i,j −1),(i,j ) = cTi,j (RTi,j −1 + RTi,j )
and
1 1 1 1 1
= + + + .
τ(i,j ),(i,j ) τ(i+1,j ),(i,j ) τ(i−1,j ),(i,j ) τ(i,j +1),(i,j ) τ(i,j −1),(i,j )
Equation (5) is generic and can be applied to any geometry by varying only the area and
the volume to be considered. The stability criterion is t
τ(i,j ),(i,j ) .
An appropriate boundary condition must be provided to define the heat input in every
portion of the domain boundary, as well as an initial condition (usually fixed temperatures:
metal pouring temperature and environment temperature at the mould cooled surface). The
solidification heat transfer model can feature a detailed treatment of interfaces. Heat flux in the
mould is related to variable heat transfer coefficients along the mould length. Below the mould,
water spray cooling systems and environment cooling extract heat from the strand surface, and
heat transfer coefficients can be related to water flow rates and surface temperatures.

2.3. Boundary conditions

The application of the solidification model to the continuous casting of square billets/blooms
was based on the following key assumptions:
(a) a two-dimensional heat transfer phenomenon was considered with heat flux admitted to
be negligible along the vertical direction (z) [20], which is a reasonable assumption due
to the high Peclet number: (z) ⇒ ∂T /∂z = 0;
(b) a control volume element was placed in a transverse slice which moves at the casting speed
through the different cooling regions of the continuous casting machine: Z = Vcasting · t,
where Vcasting [m s−1 ] is the casting speed;
(c) the billet symmetry permits that only one-quarter of the cross section be modelled for a
full thermal evolution characterization;
(d) the meniscus surface was assumed to be flat;
(e) an overall heat transfer coefficient characterizes the interfacial heat transfer between the
strand surface and the cooling water along the mould;
(f ) the surface temperature of molten metal is considered equal to the pouring temperature
and is given by Tpouring = Ttundish − 20 ˚C;
(g) at the range of temperatures where solidification occurs for metallic alloys, the physical
properties will be evaluated taking into account the amount of liquid and solid that coexists
in equilibrium at each temperature:
k = (kS − kL ) · fs + kL , (6)
c
= (cS − cL ) · fs + cL − (L ∗ dfs ), (7)
ρ = (ρS − ρL ) · fs + ρL . (8)
If fs = 0, the element is still liquid, and only the thermophysical properties of the liquid
are considered, and if fs = 1, the element is completely solid;
(h) the transient metal/mould heat transfer coefficients (hm ) used in this work are those
proposed by Toledo et al [21] and shown by equation (9).
−1
hm = (0.071 28 · e−time + 2.328 · e−time/9.5 + 0.698) · 1000 [W m−2 · K ], (9)
Continuous casting of steel billets and blooms 1075

where time [s] is the residence time along the mould length. The metal/sprays zones
and metal/free-radiation zones heat transfer coefficients (hs and hR ), respectively, are
determined by the IHCP method (inverse heat conduction problem) by using experimental
strand surface temperatures measured by colour infrared pyrometers located along the
continuous caster. The IHCP is only used to determine the nominal heat transfer
coefficients during casting;
(i) effects of mould oscillation, mould curvature, segregation and melt level fluctuation in the
mould were not considered in the knowledge base of metallurgical constraints [13];
( j) only metal/spray heat flow efficiencies are adjustable by the NNBA through water flow
rate variations in each sprays zone.

The most serious problem in the continuous casting processes of steel occurs when a high
strand surface reheat takes place in the secondary cooling zones and in the radiation zone,
mainly for low carbon steels (<0.25% carbon). This is due to the phase transformation during
cooling (liquid > liquid and delta ferrite > delta ferrite > delta ferrite and austenite ∼1400–
1485 ˚C) and the corresponding volumetric variation. This reheat can lead to considerable
internal and external defects such as rhomboidity, midway cracks close to the solidification
front, voids and porosity. Alloying elements are also important in the pursuit for defect-free
castings: Mn : S ratio >25–30 avoids crack formation inside the mushy zone, P < 0.017%
decreases the columnar zone, S < 0.015% decreases formation of FeS (Tfusion = 1200 ˚C),
Cu > 0.2% increases low melting impurities in grain boundaries, Al < 0.02% decreases
formation of AlN (900 ˚C) and Ti < 0.004% minimizes AlN formation [13].

3. Neural network based algorithm

The methodology proposed in this work features the development of the commonly used
solution in the field of optimization strategies applied to the continuous casting process, and
makes it possible to define a better strand thermal behaviour, from which the given set of
metallurgical constraints is required as a function of quality and equipment restrictions [22].
Much research has been done in this artificial intelligence field, and various algorithms have
been developed. Back-propagation networks, the most commonly used type of network, has
been used to convert text to speech, compress data, filter noise signals, predict breakout in the
continuous casting of slabs [23, 24] and predict mould taper design in the continuous casting
of billets [25].
The NNBA was developed as a nonlinear mapping of an input onto an output vector
space. This was achieved through layers of activation functions or neurons in which the input
coordinates are summed according to weighting values (W ) and bias to produce single outputs
or values. Figure 2 shows a schematic representation of the single NNBA, composed of a
number of nodes or units connected by links, employed for the improvement of the strand
thermal field in continuous casting of steel. The functionality is similar to a function taking
three parameters as input (input layer—temperatures) and producing one output (output layer—
spray water flows). The method is a feed forward network for which there is no recursiveness,
or the input vector of a specific neuron layer is formed only by the values of the preceding
layer.
The first step in building a neural network is to define the number of variable inputs (layer 0)
and the number of outputs (layer n). The complexity of the network should be sufficient to
capture the functionality of the relationship under consideration. The use of single or double
hidden layers with enough nodes can be sufficient to treat most problems. Each unit or neuron
has a set of input links from other units, a set of output links to other units, a current activation
1076 C A Santos et al

Figure 2. Schematic representation of the model based on the neural network.

Figure 3. Single activation function used in the NNBA.

level and a means of computing the activation level at next step in time, given its inputs and
weights. Figure 3 shows a typical unit or neuron. Each unit performs a simple computation:
it receives signals from its input links and computes a new activation level that it sends along
each of its output links.
The computation is split into two components: one is a linear component, called input
function (ini ), that computes the weighted sum of the unit’s input values; the second is a
nonlinear component, called activation function (F ), that transforms the weighted sum into
the final value that serves as the unit’s activation value (Si ).
The total weight input is the sum of the input activations multiplied by their respective
weights, as follows [26]:

ini = (Wi · ii ). (10)
The elementary computation step in each neuron computes the new activation value for
the neuron by applying the activation function to the result of the input function:

 
Si ← F (ini ) = F Wi · i i , (11)
i

where dissimilar models can be obtained by using different mathematical functions as the
activation function. In this work, a logistic function has been used, given by
1
= 0. (12)
1 + e−Si
If the activation function of the ith neuron in the j th layer is F (x), its outputs (Si,j ) can
be calculated from the outputs of the preceding layer (Si,j −1 ) and the corresponding weighting
(Wi,j,k ) values, where the k subscript indicates the neuron in the j − 1-th layer from which the
connection is being established [25]. According to Zietsman et al [25], Si,j can be given by:
Si,j = F (Wi · F (Wi−1 · Xi ), (13)
Continuous casting of steel billets and blooms 1077

where [X] is the input vector, [Si ] is the layer i output vector, [Wi ] is the layer i weight matrix
and F is the activation function. In this work, the results of strand surface temperatures obtained
by simulation with specific nominal input process parameters were used to feed the NNBA
determining the improved water flow rate for each spray cooling zone, which are re-submitted
to the mathematical model until desired surface temperatures in each zone are attained.
The second step in building an artificial neural network is to train the network, or better,
to train the weights using a learning algorithm applied to a set of training examples for the
task. The training set is fed through the network in this fashion until the network error reaches
a sufficiently low value. The network is then ready to be used in prediction mode.

3.1. Training of the neural network

The method used for the training of the neural network was training with a knowledge database
of solidification simulations results. During training, the network is exposed to a database
containing input (temperatures) and output (water flow rates) values grouped together (input–
output pairs). Each pair has a value for x, y and f (x, y). The input values are distributed
over a defined range, for instance, xmin –xmax for x and ymin –ymax for y. Before the training,
the weights (Wi ) of the network are set to random values. The training process includes
feeding the input–output pair and forwarding through the network one by one. During the
forward propagation step the activation functions are activated layer by layer. An activation
function first calculates a weighted sum of its outputs, producing a summation term Si . This
term is then fed to produce the final output layer of the neuron. The outputs of the output
layer are compared with the outputs suggested by the current input–output pair, and an error
is calculated. This error is propagated back through the layers to change the values of the
weights. This ensures that the network will make a better prediction the next time it is exposed
to the particular input vector. The back-propagation is probably the most employed training
heuristics and is particularly well adapted to feed forward algorithms. It is known as the delta
rule or Widrow–Hoff rule [25]. The back-propagation step is given by
Winew = α · G · (Y − Si ) + M · Wiold , (14)
where α is the learning rate, G is the inverse of activation function, Y is the desired vector and
M is the momentum, and
Winew = Wiold + Winew . (15)
The delta learning rule used in this work is a gradient descent method [27] used to
minimize the error generated by the network. The absolute values of the learning rate (α)
and the momentum (M) determine the magnitude of change made to the weights at any stage
of the training process. The α/M ratio determines the influence that new data have on weight
change relative to the influence of what had already been learned. When learning starts, the
ratio should be large. As the learning process continues, the ratio should decrease with increase
in learning. The general lines and the basic steps of the training procedure adopted in this work
are the following:
(1) initialize the parameters of the network: pre-assigned input/output pairs (Tsurface /ω) and
weight randomly (Wi,j );
(2) calculate the error e with respect to input–output pairs based on evaluating the distance
between a new object and nearest objects for which the classes they belong to are known
by the Euclidian method:

e= (W − Wi,j )2 , (16)
i
1078 C A Santos et al

(3) modify the network parameters according to a gradient descent strategy and a specific
learning rate (α = 0.5) and the momentum rate (M = 0.1). This is made by an interaction
between the current input parameters and the input–output pair.
(4) iterate from 2 to 5, modifying Wi,j until a defined error level is reached. In this work, a
maximum number of 100 000 cycles until convergence has been used.
For this procedure an auxiliary algorithm to train the neural networks was developed in
Visual Basic (Visual Basic is the trademark of Microsoft Corporation). An error level (about
1%) was assumed in relation to the final strand surface temperature and that the three networks
have identical structures. The adjustment of the first water flow rate approximates the values
of the posterior water flow rate based on surface temperatures at the end of each cooling zone.
The main input information for the determination of the ideal water flow rate for each zone is
the strand surface temperature at the beginning of each zone. This has permitted the division
of the neural network into three networks with three inputs and only one output operating in
series, improving the time of learning. After the network has been trained with the training set,
it will be able to produce valid interpolation within the input value limits. During prediction
a set of input values, within the limits xmin –xmax defined by the training set, is fed into the
network. The network processes these values in a feed-forward manner only.

3.2. Implementation of the NNBA into the mathematical model

Figure 4 shows a representation of the NNBA in the mathematical solidification model, where
the finite difference model calls the NNBA a subroutine.
The first layer has the function of normalizing the input data giving a fraction between 0 and
1, by vnorm = (v − xmin )/(xmax − xmin ), where v is a current value of the variable x, and xmin
and xmax are minimum and maximum variable values in the data set, respectively. The second
layer determines the coefficients that correlate the output as a function of input temperatures.
There are two different neurons: one representing a quality analysis and another representing a
quantity analysis. The quality analysis determines a value that will be implemented in the water
flow rate, and the quantity analysis is made by multiplying the term that correlates the input
temperatures with the water flow rate gradient. The third layer is responsible for the association
between two neurons summing their outputs and the fourth layer multiplies the found value
by the water flow rate for determining the real desired water flow rate (correction in the water
flow rate for the next search).
Seven input and three output nodes assumed in the network structure are the consequence of
the assumption that the thermal behaviour of the strand analysed is affected mainly by the water
flow rates in the secondary spray cooling zones. After preliminary tests, the network based
model was coupled to the solidification mathematical model in order to permit calculations.

4. Experimental procedure

The experimental data were obtained in an industrial caster by using infrared mobile pyrometers
located at selected positions along the secondary cooling and radiation zones. The machine
was divided into 11 different ‘regions’: mould, sprays zone I, sprays zone II, sprays zone III,
free I (free: radiation cooling zone), free II, free III, free IV, unbending point, free V, free VI
and free VII. Surface temperatures were measured at points located in the middle or end of
each region or segment, which were called ‘positions’. The temperature variations that occur
between individual rolls and spray impingement regions have been ignored.
A schematic representation of the caster machine with the positions where surface
temperatures were monitored is shown in Figure 5.
Continuous casting of steel billets and blooms 1079

Figure 4. Diagram representation of the NNBA coupled with the mathematical solidification
model.

In this study experiments were conducted with both a low carbon steel (SAE 1007) and
a medium carbon steel (SAE 1025). The input parameters used in casting operations and
simulations, as well as the thermophysical properties are presented in table 1.
The temperature files containing the experimentally monitored temperatures during
continuous casting operations were used by a finite difference heat flow program (inverse
heat conduction solution) in order to determine nominal transient metal/sprays and
metal/environment heat transfer coefficients (hS and hR ), respectively, as described in earlier
papers [28, 29]. Spray/ingot heat transfer coefficients used in this work are based on an
experimental equation proposed in the literature [30], and they are related to water flow rates
along the different sprays zones along the machine by
 o 0.556
hspray = 0.36 · ω · 1000, (17)
o
where hspray [W m−2 ·K−1 ] and ω is the water flow rate at each zone [l s−1 ].

5. Results and discussion

5.1. Nominal conditions

To determine the metal/spray heat transfer coefficients for each spray cooling zone as a
function of water flow rates and heat transfer coefficients for each free zone as a function
of surface temperatures the experimental strand surface temperatures were used. These values
are presented in tables 2 and 3 for both SAE 1007 and the SAE 1025 steels, respectively.
1080 C A Santos et al

Figure 5. Positions where surface temperatures were measured along the continuous caster.

Table 1. Input parameters for the SAE 1007 billet and the SAE 1025 bloom used in simulations
performed with the mathematical solidification heat transfer model.

Units Billet Bloom

Dimension mm 150 × 150 240 × 240
Mould length mm 685 685
Metal 1 007 steel 1 025 steel
Specific heat J kg −1 K−1 cS : 600 cL : 575 cS : 640 cL : 600
Density kg m−3 ρS : 7800 ρL : 7000 ρS : 7 850 ρL : 7 000
Thermal conductivity W m −1 K−1 kS : 46.0 kL : 30.0 kS : 30.7 kL : 31.0
Latent heat of fusion J kg−1 260 000 277 000
Solidus temperature ˚C 1 449 1 468
Liquidus temperature ˚C 1 514 1 504
Sprays 1 (length/water flow) (mm)/(l s−1 ) 1 926 1.50 1 926 0.59
Sprays 2 (length/water flow) (mm)/(l s−1 ) 855 0.60 855 0.25
Sprays 3 (length/water flow) (mm)/(l s−1 ) 1 045 0.80 1 045 0.20
Casting rate m min−1 2.04 0.71
Melt temperature in tundish ˚C 1 558 1 545
Machine length m 24.0 24.0
Continuous casting of steel billets and blooms 1081

Table 2. Experimental strand surface temperatures measured during casting: 1007 steel.
Positions Measured temperatures (average ˚C)
4 1060
5 1134
7 1181
8 1251
9 1290
10 1231
11 1272
Section (mm) Ttundish (˚C) Tpouring (˚C) TLiquidus (˚C) TSolidus (˚C) Vcasting (m min−1 )
150 × 150 1558 1538 1514 1449 2.04

Table 3. Experimental surface temperatures measured during casting: 1025 steel.

Positions Measured temperatures (average ˚C)
4 1039
5 1039
7 1061
8 1152
9 1167
10 1181
Section (mm) Ttundish (˚C) Tpouring (˚C) TLiquidus (˚C) TSolidus (˚C) Vcasting (m min−1 )
240 × 240 1545 1525 1504 1468 0.71

5.1.1. SAE 1007. Figure 6 shows a comparison between experimental and nominal simulated
results for the 1007 steel billet. The simulation was based on metal/mould heat transfer
coefficients proposed by Toledo et al [21], and the metal/sprays and metal/environment heat
transfer coefficients were determined by the IHCP method utilizing the temperatures shown
in table 2. A good agreement can be observed with differences between the experimental and
the simulated being lower than 15 ˚C, as shown in table 4.
The main metallurgical parameters calculated by the mathematical model were: solid shell
thickness at the mould exit: 13 mm; strand surface temperature at the mould exit: 1181 ˚C;
strand surface temperature at the unbending point: 1143 ˚C; strand surface temperature at the
flame cut-off point: 1070 ˚C and point of complete solidification: 16.93 m from the meniscus.
As observed in figure 6, the nominal condition imposes maximum reheat between mould
exit and free zones of about 116 ˚C and reheat of about 63 ˚C between the last sprays zone
and the free cooling zone. The excessive reheat between mould exit and spray I (about
88 ˚C) can develop defects such as surface, longitudinal, midway and transverse cracks. The
recommended maximum reheating of the strand surface temperature is <100 ˚C [5]. The strand
surface temperature profiles along the three different sprays zones exhibit thermal gradients
of about 94 ˚C. For low carbon steels (<0.185 ˚C), the temperature range conducive to low
ductility and poor mechanical properties are: high temperature zone (Tsolidus − 40 ˚C < T <
Tsolidus ) due to P and S contents; intermediate temperature zone (A3 < T < 1100 ˚C) due to the
Mn : S rate and the presence of carbides and nitrides in grain boundaries and low temperature
zone (650 < T < 750 ˚C) due to AlN, carbides and nitrides in grain boundaries. The strand
surface temperature at the unbending point is far from the intermediate temperature zone of low
ductility. The point of complete solidification (16.93 m) is not very far from the unbending point
(16.4 m) and can cause defects such as external, internal and central cracks. If the temperature
at the centre of the strand is lower than 1350 ˚C (high temperature zone of low ductility), pinch
rolls cracks can be prevented. In this case, the temperature at the centre is about 1233 ˚C.
1082 C A Santos et al

Table 4. Correlation between water flow rates and metal/sprays and metal/environment heat transfer
coefficients, and measured and nominal calculated surface temperatures (1007 steel).

Heat transfer Nominal

Nominal coefficients Experimental calculated
Regions flow (l s−1 ) (W m−2 ·K−1 ) temperatures (˚C) temperatures (˚C)
Sprays I 1.50 450 — —
Sprays IIa 0.60 270 End 1272 1275
Sprays IIb 0.80 315 Middle 1231 1258
Free I 125 Middle 1290 1296
Free II 185 Middle 1251 1259
Free III 285 End 1181 1189
Free IV 305 End 1134 1143
Unbending 310 End 1060 1171
Free V 200 — —
Free VI 175 — —
Free VII 215 — —

1600
1550 T p o u rin g S u rfa c e T e m p e ra tu re
1500 S A E 1 0 0 7 S te e l B ille t
1450
1400
In d u s trial d a t a
1350
P 9: 1290 S im u la te d
Temperature [ C]

1300 P 11: 1272

o

P 8: 1251
1250 P : p o s itio n w h e re e xp e rim e n ta l
s u rfa c e te m p e ra tu re w a s m e a s u re d
1200 P 10: 1231

1150 P 7: 1181

1100 P 5: 1134

1050 P 4: 1060
1000
Unbending Point

950
Sprays III

900
Sprays II

Free VII
Sprays I

Free VI
Free IV

Free V
Free III
Free II
Mould

Free I

850
800
0 2 4 6 8 10 12 14 16 18 20 2 4
Position from the meniscus [m]

Figure 6. Comparison between nominal calculated and experimental strand surface temperatures:
1007 steel.

Figure 7 shows a comparison between experimental and nominal calculated strand surface
temperatures.

5.1.2. SAE 1025. The main metallurgical parameters calculated by the mathematical model
were: solid shell thickness at the mould exit: 25 mm; strand surface temperature at the mould
exit: 1038 ˚C; strand surface temperature at the flame cut-off point: 845 ˚C and point of complete
solidification: 14.34 m from the meniscus. A good agreement can be observed with differences
between experimental and nominal simulated being lower than 15 ˚C, as shown in table 5. Even
so, the strand surface temperature reheat between mould exit, sprays zones and free zones are
Continuous casting of steel billets and blooms 1083

1600
1550 T p o u rin g
S u rfa c e T e m p e ra tu re
1500 S A E 1 0 2 5 S te e l B lo o m
1450
1400
1350
In d u s tria l d a ta
S im u la te d
Temperature [ C]

1300
o

1250
1200 P 10: 1181
P 8: 1152
1150 P 9: 1167
1100 P 7: 1061
P 5: 1039
1050
1000

Unbending Point
P 4: 946 P 2: 878
950 P 3: 917 P 1: 858
Sprays III

900
Sprays II

Free VII
Sprays I

Free VI
Free IV

Free V
Free III
Free II
Mould

Free I

850
800
0 2 4 6 8 10 12 14 16 18 20 2 4
Position from meniscus [m]

Figure 7. Comparison between nominal calculated and experimental strand surface temperatures:
1025 steel.

Table 5. Correlation between water flow rates and metal/spray and metal/environment heat transfer
coefficients, and measured and nominal calculated surface temperatures (1025 steel).

Heat transfer Nominal

Nominal coefficients Experimental calculated
Locations flow (l s−1 ) (W m−2 ·K) temperatures (˚C) temperatures (˚C)
Spray I 0.59 270 — —
Spray IIa 0.25 150 — 1185
Spray IIb 0.20 166 Middle 1181 1175
Free I 155 Middle 1167 1169
Free II 166 Middle 1152 1138
Free III 200 End 1061 1041
Free IV 185 End 1039 1033
Unbending 190 End 946 943
Free V 190 End 917 915
Free VI 195 End 878 870
Free VII 166 End 858 849

quite high (for instance, about 125 ˚C in spray I). Due to the strand dimensions and casting
speed, the point of complete solidification is anticipated (about 2.15 m before the unbending
point: 16.5 m), which could lead to an excessive stress level at the unbending point. In both
cases, the observed differences between measured and nominal simulated values are mainly
due to non-variable heat transfer coefficients assumed along each sprays zone.

5.2. Training of the neural network

After determining the metal/spray heat transfer coefficients for each cooling zone by the IHCP
method, these parameters were used in the training of the NNBA. The values in the sprays
1084 C A Santos et al

T1 T2 T3 T4 T5 T6 T7

ωmax
Nominal
water flow ωmin
rate
∆ω1 ∆ω2 ∆ω3

Sprays Zone I Sprays Zone II Sprays Zone III

P1 ω1 P3 ω2 P5 ω3 P7

Ti : temperature at selected positions

ωi : water flow rate for each spray zone
∆ωi : water flow rate gradient
Pi : selected positions at beginning, middle and end of each zone

Figure 8. Representation of the strategy applied to develop correlations between surface

temperature and water flow rate (input–output pairs).

zones were converted in terms of water flow rates according to the formulation proposed
by Bolle and Moureau [30]. A set of simulations with the mathematical solidification heat
transfer model and nominal spray cooling data were performed to build correlations between
the strand surface temperature for each zone and the water flow rate, Tsurface = f (ω) and
consequently to obtain input–output pairs. The training input parameters used for this set of
network were: learning rate 0.5 and momentum 0.4. The entire training set was presented to
the network several times to ensure that the influence of all input–output pairs (Tsurface , ω) was
incorporated into the weight values (W ).
Figure 8 shows a schematic representation of the search strategy used in this case. For
the training, the algorithms run with limits of iterations setup to 10 000, 25 000 and 100 000
iterations in a maximum period of about 5 min. In this figure, ωmin and ωmax represent minimum
and maximum water flow rates at caster sprays zones, and the increment represents the common
variation in water flow rates adopted in industrial practice.

5.3. Improving cooling conditions

The results of simulations are presented in figures 9 and 10 for the 1007 steel billet and the
1025 steel bloom, respectively. Simulated results (both nominal and NNBA) and industrial
measurements of surface temperatures for the 1007 steel billet are compared in figure 9.
As shown in figure 9, a more homogeneous surface temperature profile along the sprays
zones is attained with the NNBA, with lower thermal gradients between adjacent cooling
sprays zones (maximum reheat: 80 ˚C in sprays zone I and 15 ˚C in sprays zones II and III).
The algorithm suggests an increase in the water flow rate for zones I, II and III, permitting
the decrease of strand reheat along the secondary spray-cooling and free radiation zones. The
desired strand surface temperatures at the middle of each sprays zone have been assumed as
follows: T2 : 1250 ˚C (zone I) and T4 and T6 : 1225 ˚C (zone II and zone III), and the simulations
based on the neural network have found these expected values. These temperatures were chosen
based on the steel chemical composition (low carbon steel), casting speed (high speed) and
recommended industrial practice. In this case, the final of solidification occurred at 16.44 m
from the meniscus, satisfying the metallurgical criterion which dictates that solidification must
be complete before the unbending point (rather than 16.93 m for the nominal condition). The
surface strand temperature at the unbending point was found to be 1150 ˚C and the temperature
at the flame cut-off was found to be 1051 ˚C, which are lower than the values which resulted
Continuous casting of steel billets and blooms 1085

1600
1550
T pouring SAE 1007 Steel Billet
1500
1450
1400
1350 Industrial
Temperature [ C]

1300 Nom inal

o

NNBA - surface
1250
NNBA - centre
1200
1150
1100
1050
1000

Unbending Point
950
Sprays III

900
Sprays II
Sprays I

Free VII
Free IV

Free VI
Free III

Free V
Free II
Mould

Free I

850
800
0 2 4 6 8 10 12 14 16 18 20 2 4
Position from m eniscus [m ]

Figure 9. Comparison of results of nominal and improved NNBA strand surface temperature
during continuous casting of the 1007 steel billet.

1600
T pouring SAE 1025 Steel Bloom
1500

1400
Temperature [ C]

1300 Industrial
o

Nom inal
1200 NNBA - surface
NNBA - centre

1100

1000
Unbending Point
Sprays III

900
Sprays II
Sprays I

Free IV

Free VI
Free III
Free II

Free V
Mould

Free I

800
0 2 4 6 8 10 12 14 16 18 20 2 4
Position from meniscus [m]
Figure 10. Comparison of results of nominal and improved NNBA strand surface temperatures
during continuous casting of the 1025 steel bloom.
1086 C A Santos et al

Table 6. Water flow rates for nominal and improved operational conditions.

Steel Zones Nominal (l s−1 ) Improved (l s−1 ) Nominal (˚C) Improved (˚C)
1007 I 1.50 1.560 Middle 1268 end 1233 1250 1209
II 0.60 0.700 Middle 1255 end 1275 1225 1221
III 0.80 0.840 Middle 1258 end 1248 1225 1219
1025 I 0.59 0.712 Middle 1170 end 1130 1140 1099
II 0.20 0.299 Middle 1181 end 1185 1140 1143
III 0.25 0.309 Middle 1175 end 1160 1140 1136

for the nominal condition. It can be seen that for the nominal condition, the reheat between the
mould exit and free I region is still high, however it has decreased from 116 ˚C for the nominal
condition to 85 ˚C for the improved condition.
For the 1025 steel, the assumed ideal strand surface temperature at the middle of each
sprays zones was: sprays zone I, II and III: 1140 ˚C. The NNBA suggests an increase in water
flow rate for zones I, II and III. As expected, the increase in water flow of sprays zones II and
III in the search of ideal temperatures has induced a decrease of strand surface temperatures
of subsequent zones, as shown in figure 10.
The strand surface temperature along sprays zones II and III has experimented a continuous
decrease anticipating the point of complete solidification when compared with the nominal
simulated conditions. The new main metallurgical parameters calculated by the mathematical
model were: solid shell thickness at the mould exit: 25 mm; strand surface temperature at
the mould exit: 1038 ˚C; surface temperature reheat in sprays zone I: 88 ˚C; strand surface
temperature at the flame cut-off point: 849 ˚C; and point of complete solidification: 13.70 m
from the meniscus.
Table 6 indicates nominal and improved water flow rates in each sprays zone for both
steel grades, as well as the nominal and calculated temperatures at the middle and at the end
of each sprays zone. As shown in table 6, the differences between expected strand surface
temperatures at the middle of each sprays zone (T2 , T4 , T6 , see figure 8) and improved strand
surface temperatures at the same positions were lower than 1 ˚C.
As expected, the neural network algorithm has found the best strand surface temperatures
at specific locations. Future works will deal with a more accurate determination of heat transfer
coefficients along each spray cooling zone, permitting the insertion of relationships such as
hsi = f (ωi .Pi ) into the model, where Pi is the position along each sprays zone.

6. Conclusions

Heat transfer coefficients have been determined from surface temperature industrial
measurements concerning billet and bloom continuous casting operation as a pre-requisite
for improving simulations. Infrared mobile pyrometers have been successfully employed
to measure strand surface temperatures along the machine during the continuous casting
of steel billet (SAE 1007) and bloom (SAE 1025). Model predictions of strand surface
temperature profiles have been compared to measured profiles, and the agreement is quite
good. The solidification mathematical model working integrated with a NNBA has permitted
the improvement of cooling conditions (secondary cooling) for the continuous casting of steel,
i.e. the search method permits a more homogeneous strand surface temperature profile to be
attained and with lower thermal gradients between adjacent sprays zones. It is important to
point out that more accurate simulations can be achieved if particular heat transfer formulations
are developed for a specific continuous caster by using approaches such as a fitting technique
 Continuous casting of steel billets and blooms 1087

between theoretical–experimental thermal profiles and if transient heat transfer coefficients

along sprays and free-radiation zones are determined.

Acknowledgments

The authors acknowledge financial support provided by FAPESP (The Scientific Research
Foundation of the State of São Paulo, Brazil), CNPq (The Brazilian Research Council) and the
steel company, Aços Finos Piratini—Gerdau. Tiago Ferraz is also acknowledged for helping
with the computer programming.

References

[1] Lally B, Biegler L T and Henein H 1991 Metall. Trans. B 22 641–8

[2] Lally B, Biegler L T and Henein H 1991 Metall. Trans. B 22 649–59
[3] Kumar S, Walker B N, Samarasekera I V and Brimacombe J K 1993 Proc. 13th PTD Conf. (Nashville, TN)
pp 119–41
[4] Kumar S, Meech J A, Samarasekera I V and Brimacombe J K 1993 Iron Steelmaker 20 29–36
[5] Samarasekera I V, Brimacombe J K and Wilder K 1994 Iron Steelmaker 21 53–63
[6] Spim J A, Santos C A, Ierardi M C F and Garcia A 1997 Proc. 4th Decennial Int. Conf. on Solidification
Processing (Sheffield, England, 1997) pp 166–9
[7] Filipic B and Sarler B 1997 Proc. 2nd Int. Metallurgical Conf. on Continuous Casting of Billets (Trinec, Czech
Republic, 1997) pp 161–8
[8] Filipic B and Sarler B 1998 Proc. 6th European Congress on Intelligent Techniques and Soft Computing:
EUFIT’98 (Aachen, Germany) vol 1, pp 444–9
[9] Brimacombe J K 1993 Iron Steelmaker 21 35–47
[10] Brimacombe J K 1999 Metall. Mater. Trans. A 30 1899–912
[11] Cheung N and Garcia A 2001 Eng. Appl. Artif. Intell. 14 229–38
[12] Santos C A, Spim J A, Ierardi M C F and Garcia A 2002 Appl. Math. Model. 26 1077–92
[13] Santos C A, Spim J A and Garcia A 2003 Eng. Appl. Artif. Intell. 16 511–17
[14] Chakraborti N, Kumar R S and Jain D 2001 Appl. Math. Model. 25 287–97
[15] Chakraborti N, Kumar R S and Roy G G 2003 J. Mater. Eng. Perform. 12 430–5
[16] Welty J R 1976 Engineering Heat Transfer (New York: Wiley)
[17] Smith G D 1985 Numerical Solutions of Partial Differential Equations 3rd edn (Oxford: Clarendon)
[18] Poirier D R and Poirier E J 1994 Heat Transfer Fundamentals for Metal Casting (Warrendale, PA: The Minerals,
Metals and Materials Society)
[19] Spim J A and Garcia A 2000 Numer. Heat Trans. B 38 75–92
[20] Brimacombe J K, Samarasekera I V and Lait J E 1984 Iron and Steel Society of AIME vol 20
[21] Toledo G A, Lainez J and Cirión J C 1993 Mater. Sci. Eng. A 173 287–91
[22] Bhadeshia H K D H 1999 ISIJ International 39 966–79
[23] Kominami H 1991 Nippon Steel Technical Report No 49
[24] Sowka E, Dikhoff P S, Harder J, Munscher F and Beirer G 1999 Iron Steel Eng. 30–5
[25] Zietsman J H, Kumar S, Meech J A, Samarasekera I V and Brimacombe J K 1998 Ironmak. Steelmak. 25 476–83
[26] Russel S J and Norvig P 1998 Artificial Intelligence: A Modern Approach (New Jersey: Prentice Hall)
[27] Crivelaro K C O, Seleghim P and Hervieu E 2002 J. Brazilian Soc. Mech. Sci. 24 70–75
[28] Santos C A, Quaresma J M V and Garcia A 2001 J. Alloys Compounds 319 174–86
[29] Raudenský M, Horský J, Krejsa J and Sláma L 1996 Proc. COST 512, General Workshop (Davos, 1996)
ed M Rappaz and M Kedro, pp 203–8
[30] Bolle E and Moureau J C 1979 Proc. Two Phase Flows and Heat Transfer III—NATO Advanced Study Institute
(Dubrovnik, Yugoslavia) 1327–46

View publication stats