MATERIALS AND MANUFACTURING PROCESSES

Vol. 18, No. 3, pp. 401–408, 2003

Process Control and Optimization of the AOD Process

Using Genetic Algorithm
Downloaded by [University Library Utrecht] at 10:40 13 November 2012

Brahma Deo* and Vivek Srivastava

Department of Materials and Metallurgical Engineering,

Indian Institute of Technology, Kanpur, India

ABSTRACT

The AOD process of making stainless steel is very popular because of its fast rate
of decarburization and excellent control of composition and temperature. Both
static and dynamic models are available for process control. The objective of
the present study is to demonstrate the applicability of GA in adapting a process
control model to a particular plant environment and subsequently optimize the
blowing scheme to reduce overall costs. Deviation of the predicted results of
the process control model, from the actual plant data, is first minimized by adjust-
ing the chemical reaction rate parameters and heat balance equation. It is found
that the adaptation of the process control model to an actual situation is feasible
only below 0.8% carbon. After adapting the model, optimization is done through
GA wherein the objective function contains the cost of refractory, process time,
chromium loss etc. It is thus possible to devise a blowing scheme to arrive at
the desired end point composition and temperature at the lowest cost.

*Correspondence: Brahma Deo, Department of Materials and Metallurgical Engineering,

Indian Institute of Technology, Kanpur 208016, India; E-mail: [email protected].

401

DOI: 10.1081/AMP-120022018 1042-6914 (Print); 1532-2475 (Online)

Copyright # 2003 by Marcel Dekker, Inc. www.dekker.com
 402 Deo and Srivastava

1.0 INTRODUCTION

The Argon Oxygen Decarburization (AOD) process of stainless steel making is

very popular for reasons of economy and efficient process control. It allows the use
of cheap high carbon ferro-chrome instead of costly low carbon ferro-chrome as a
source of chromium. The efficiency and productivity of the AOD process is very high
because of a relatively short process time, due primarily to the fact that an argon-
oxygen (or air -oxygen) mixture is injected at the bottom of the vessel through spe-
cially designed copper tuyeres which enhance the rate of carbon oxidation. A brief
description of the process is given below so as to explain the optimization problem
Downloaded by [University Library Utrecht] at 10:40 13 November 2012

and for more details of the thermodynamics and kinetics of reactions, reference may
be made to literature.[1–14] To the best of our knowledge, this is the first time that the
process control model of the AOD process at a mini steel plant has been optimized
by using the Genetic Algorithm (GA). Similar examples of the application of GA to
optimize of other metallurgical processes can be found in literature.[15–18]
A schematic diagram of an 8-ton AOD vessel studied in the present work is
shown in Fig. 1. The height of the vessel is approximately 2.7 m. The diameter of
the cylindrical portion is approximately 1.4 m. The free internal volume is 4.15 m3.
The vessel is mounted on trunions and can be tilted for charging and tapping.
The liquid semi-product, which is the initial charge for blowing in AOD, is pre-
melted in electric arc furnaces starting from a charge of cheap high carbon-ferro
alloys. The semi-product (containing, say 2% carbon, 22% chromium, 2.15% nickel
and balance iron and slag) is poured into the AOD converter. The blowing is done in
three stages. For a typical heat, the operation proceeds as follows.
Stage 1: The first stage of the blow lasts for roughly 25 min. The flow rates of
oxygen and air used are 5.0 m3=min and 0.6 m3=min, respectively. The melt contains
1.00% carbon and 20% chromium at the end of the first stage of blow. The tempera-
ture of the melt is around 1650
C.

Figure 1. A typical AOD vessel (8 tons).

 Process Control and Optimization of AOD Process 403

Stage 2: The second stage of the blow is continued for around 35 min. The flow
rates of oxygen and air are 5.0 m3=min and 1 m3=min, respectively. During the blow,
lime and high carbon ferro-alloys are added to bring down the temperature and con-
trol the metal and slag chemistry. At the end of the blow the melt contains about
0.2% carbon, 18% chromium, while the temperature rises to around 1700
C.
Stage 3: The third stage of the blow is continued till the target carbon content is
achieved. The flow rates of oxygen and air are about 1.75 m3=min and 3 m3=min,
respectively. During this stage, scrap and ferro alloys may be added as coolant into
the melt. The temperature of the melt at the end of this blow is approximately
1675
C.
Downloaded by [University Library Utrecht] at 10:40 13 November 2012

Stage 4: In this stage, alloying additions of nickel and boron are made. Finally,
ferro-silicon is added to reduce chromium oxide in the slag. During the reduction
period, argon and air are blown to ensure good mixing and reduction. Then, the slag
off is done followed by tapping of the stainless steel into a ladle.
Several trials conducted in the beginning of the present work showed that stage 1
is always very unpredictable, hence only stage 2 and stage 3 are optimized in this
work. The last stage is relatively easy to control as it essentially involves recovery
of oxidized chromium from slag by reduction with ferro-silicon under conditions of
gentle inert gas stirring.

2.0 GOVERNING EQUATIONS FOR OXIDATION OF CARBON,

CHROMIUM AND RISE OF BATH TEMPERATURE

It is known that above 9% Cr, decarburization is controlled by Cr3O4 present in

slag. Suppose dQd m3=ton and dQo m3=ton are the amounts of diluent gas and oxygen
injected into the melt in time interval dt at a time t. The amount of carbon oxidized,
dCt,in this time interval, dt, will be

aðdQd ÞK½Ct
d½Ct ¼ h i ð1Þ
18:7 ½Cr3=4 Ptot  Kfc ½C t

where Ct and [Cr]t are carbon and chromium contents at time t, respectively. K is the
equilibrium constant for the chemical reaction

1=4ðCr3 O4 Þs þ ½C ¼ 3=4½Cr þ ðCOÞg ð2Þ

for which the equilibrium constant, K, at temperature, T, is given by

log K ¼ 11520=T þ 7:64 ð3Þ

The Henrian activity coefficient, fc, depends upon bath composition and temperature
at any time. The adjustment factor a is introduced in Eq. (1) to take into account the
non-ideal behavior of the slag-metal-gas reaction system. The oxygen, that is left
over after the decarburization reaction, is used in oxidation of chromium and the
amount of chromium oxidized, dCrt, is given by
 404 Deo and Srivastava

dCrt ¼ b  0:348½dQo  9:35  dCt  ð4Þ

Similar to the factor a in Eq. (1), the factor b in Eq. (4) is introduced to take care of
non-ideal behavior of the system. For example, according to reaction (2), all the
Cr3O4 formed initially due to oxidation of chromium may not be subsequently
reduced by the carbon in the melt due to mass transport limitations.
The temperature rise of the bath, dT, is predicted by heat balance calculations
during each time step1

dHC þ dHCr þ dHFe þ dHd

dT ¼ þ Vt  dt ð5Þ
m  Cp
Downloaded by [University Library Utrecht] at 10:40 13 November 2012

dHC ¼ 57.754 T  156818.88 J=mol, heat of oxidation of carbon,

dHCr ¼ 169.443T  1225476.8 J=mol, heat of oxidation of chromium,
dHFe ¼ 134.23T  317278.08 J=mol, heat of oxidation of iron,
dHd ¼ r  dQd  Cp  (Tout  Tin ) J=mol, sensible heat of diluent gas,

where r is the density of air (kg=m3), Tout is the temperature of outgoing gas (K), Tin
is the temperature of incoming gas (K) and Vt is the rate of cooling in K=min, which
can be estimated from

Vt ¼ k½0:0911  ðT  473Þ þ 2:73  1010 ðT4  2984 Þ=ðm  CpÞ ð6Þ

where m is the mass of the melt (kg), Cp is the heat capacity of the melt (kJ=kg=K)
and k is an empirical parameter similar to a and b to account for non-ideal behavior.

3.0 OPTIMIZATION OF PROCESS PARAMETERS BY USING

THE GENETIC ALGORITHM

The process control model is based on the recursive solution of Eqs. (1)–(6) at
each time step to predict carbon, chromium and temperature at the end of stage 2
and stage 3. The values of a in Eq. (1), b in Eq. (4) and k in Eq. (6) are optimized
by using GA in the present work. The objective function is written such that all the
three terms, i.e., C, Cr and T get equal weighting, due to the square penalty function.

f ¼ Sf½10ðCactual  Cpre Þ2 þ ½ðCractual  Crpre Þ2 þ ½0:1ðTactual  Tpre Þ2 g

where the values of Cactual, Cractual and Tactual are obtained from actual plant data
and Cpre, Crpre and Tpre are obtained from recursive solution of Eq. (1), Eq. (4)
and Eq. (6), respectively, for small time intervals dt integrated over the entire per-
iod of a stage. It can be seen that the objective function is complex with multiple
possibilities of solution. GA is ideally suited to such a situation. The selection
procedure used is that of a Roulette wheel probabilistic generation. The structures
are treated as rings, choosing two crossover points and exchanging the section
 Process Control and Optimization of AOD Process 405

Table 1. Optimized process parameters for stage 2 and stage 3.

a b k

Stage 2 0.985 1.31 1.04

Stage 3 0.972 1.41 1.14

between these points. The initial population for this problem was taken as 100 and
the gene length was 30 (i.e., 10 3). The optimal values of the process parameters
determined from several heats made in the plant are shown in Table 1.
The optimum a value, as expected, decreases from stage 2 to stage 3 due to a
Downloaded by [University Library Utrecht] at 10:40 13 November 2012

decrease in the carbon content of the bath, and for the same reason b and k values
increase due to excessive chromium oxidation. The value of b (obtained as greater
than unity) essentially indicates that significant oxidation of chromium occurs due
to direct oxygen pickup from the atmosphere. When the optimal values of a, b
and k are used for the AOD process (see typical results for a heat in bar charts in
Fig. 2) a very close fit with actual values of carbon, chromium and temperature is
obtained in both the stages.

4.0 COST OPTIMIZATION

The process model developed above for the prediction of carbon, chromium and
temperature can now be coupled with costs of refractory, gases consumed during
blowing, chromium loss in slag, fixed cost and a temperature penalty for deviations
from the target temperature, so as to ensure overall minimum cost of operations. The
objective function used in the present study to minimize overall cost is

f ¼ 0:483  F  fCr%initial  massinitial  Cr%final  massfinal g=100

þ 0:24  F  ½fT  ð1650
þ deviationÞg2 ShT  ð1650
þ deviationidt
þ fð1650
deviationÞTg2 Shð1650
deviationÞ  Tidt
þ 1:0  F  ttot þ 0:024  F  fT  1700
g þ 1:0  F  104  Qair
þ 5:0  F  104  QO2

where ‘deviation’ is the range over which the temperature is allowed to vary,

hai ¼ a if a > 0
hai ¼ 0 if a < 0

and F is fixed cost at the plant.

Since the cost optimization needs to be carried out for every blow, it is necessary
to ensure the fast convergence of the GA. Therefore, average fitness value of each
generation versus crossover and mutation probability were studied and the fastest
convergence was obtained for crossover probability of 0.8 and mutation probability
of 0.01. These values were, hereafter, kept constant.
 406 Deo and Srivastava
Downloaded by [University Library Utrecht] at 10:40 13 November 2012

Figure 2. Bar charts showing comparative values of carbon, chromium and temperature for
typical heats.

The parameter ‘deviation’ is introduced in the objective function to see the effect
of admissible temperature variation at the end of each stage on the process cost. To
start with, optimization runs were carried out to find the optimal oxygen flow rate
for a given air flow rate assuming a fixed temperature deviation of 5 K. The results
are shown in Fig 3. Air being inexpensive material, function value decreases with
increasing air flow rate.
The effect of ‘deviation’ on process cost at maximum air flow rate of 5 m3=min is
shown in Table 2. The higher the admissible deviation the lower is the cost,
but beyond 10 K, the oxygen flow rate stabilizes around 2.72 m3=min for the AOD
converter studied in the present work. This value is dependent upon converter
geometry and nozzle design chosen in the present work and will thus vary from
 Process Control and Optimization of AOD Process 407
Downloaded by [University Library Utrecht] at 10:40 13 November 2012

Figure 3. Optimal oxygen flow rate for a given maximum air flow rate.

Table 2. Optimal flow rate of oxygen and relative cost function value at different values
of deviation.

Deviation Air flow rate O2 flow rate Relative cost

(K) (m3=min) (m3=min) function value

5 5.0 2.5 1.010

10 5.0 2.72 1.005
15 5.0 2.72 1.000

one converter to another. As a particular case if the deviation allowed in the tem-
perature is ±5 K, then the savings due to optimization would be approximately 1%
of the total cost.
The significant result of the cost optimization model is that for a given initial
condition in each stage, we can determine the optimal oxygen flow rate, provided
the maximum permissible air flow rate is known, which, in turn, is dictated by design
of the tuyeres.

5.0 CONCLUSIONS

1. Process control model of the AOD process at a mini steel plant has been
optimized by using the Genetic Algorithm to predict chromium (<3.3%
error), carbon (<5.5% error) and temperature (<1.10% error), at the end
of stage 2 and stage 3.
2. The process control model is further integrated with the cost model to deter-
mine optimal oxygen flow rate for a given maximum permissible air flow
rate.
3. The above findings have been successfully applied in the operations of a mini
steel plant. Savings realized due to optimization increase (0.5–1%) when the
requirement of temperature control in particular heat becomes more and
more stringent.
 408 Deo and Srivastava

REFERENCES

1. Semin, A.E. Mathematical modeling of argon-oxygen decarburization of steel.

Steel in USSR 1983, 13, 95–97.
2. Hilty, D.C. Stainless steel making (Chapter 13). In Electric Furnace Steelmaking;
Taylor, C.R., Ed.; The Iron and Steel Society of AIME: USA, 1985; 143–157.
3. Leach, J.C.C.; Rodgers, A.; Sheehan, G. Operation of AOD process in BSC.
Iron making and Steelmaking 1978, 5, 107–120.
4. Fruehan, R.J. Reaction model for the AOD process. Ironmaking and Steel-
making 1977, 4, 153–158.
Downloaded by [University Library Utrecht] at 10:40 13 November 2012

5. Deb Roy, T.; Robertson, D.G.C. Mathematical model for stainless steelmaking
Parts 1 and 2. Ironmaking and Steelmaking 1978, 5, 198–205, 207–210.
6. Hodge, A.L. AOD process and its eighty-six months of growth. Ironmaking
and Steelmaking 1977, 4, 81–87.
7. Ikeda, T.; Taga, M.; Matsuo, T.; Masuda, S. Some thermochemical discussions on
oxidizing refining of steel. Trans. Iron and Steel Institute of Japan 1981, 21, 433–445.
8. Saccomano, J.M.; Choulet, R.J.; Ellis, J.D. Making stainless steel in the argon-
oxygen reactor at Joslyn. J. Metals 1969, 21, 59–64.
9. Healy, G.W.; Hilty, D.C. Effect of oxygen input rates in the decarburization of
chromium steel. J. Metals 1957, 9, 695–707.
10. Sevnic, N.; Elliot, J.F. Kinetics of reduction of Cr2O3(s) by liquid Fe-Cr-C
alloys. Ironmaking and Steelmaking 1976, 3, 268–278.
11. Tilliander, A.; Jönson, P.; Jönson, L. Effect of Ar=O2 mixtures on heat transfer
and fluid flow characteristics in AOD nozzles. Steel Research 2000, 71, 429–434.
12. Fabritius, T.; Kupari, P.; Härkki, J. Physical modeling of a sidewall blowing
converter. Scandanavian J. Metallurgy 2001, 31, 57–64.
13. Forster, D.; Mackenzie, K.; Beeley, J.; Macvinnie, R. Computerization of AOD
steelmaking at British Steel Stainless. In 11th PTD Conference Proceedings on Ultra
Refining Processes for Specialty Steelmaking, ISS AIME: Atlanta, 1992; pp. 3–8.
14. Wijk, O., Modeling of stainless steelmaking in converters. In 11th PTD Confer-
ence Proceedings on Ultra Refining Processes for Specialty Steelmaking, ISS
AIME: Atlanta, 1992, pp. 23–36.
15. Rastogi, R.; Deb, K.; Deo, B.; Boom, R. Genetic adaptive search model of hot
metal desulphurization. Steel Research 1994, 65, 472–478.
16. Deo, B.; Deb, K.; Jha, S.; Sudhakar, V.; Sridhar, N.V. Optimal operating con-
ditions of the primary end of and integrated steel plant: genetic adaptive search
versus classical techniques. Iron and Steel Inst. Jpn. Int. 1998, 38, 98–105.
17. Sheel, P.; Deo, B. Prediction of silicon content of hot metal using fuzzy-GA
regression. In Genetic Evolutionary Computation Conference (GECCO),
July 13–17; AAAI: Orlando, 1999; pp. 1685–1690.
18. Deo, B.; Datta, B.; Kukreja, B.; Deb, K.; Rastogi, R. Optimization of back
propagation algorithm and genetic adaptive search assisted ANN models. Steel
Research 1994, 65, 528–533.