Prediction of silicon content of alloy in
ferrochrome smelting using data-driven
models
by A.V. Cherkaev1, M. Erwee1, Q.G. Reynolds2,3, and S. Swanepoel1
Affiliation:
1Samancor, South Africa.
2Pyrometallurgy Division, Mintek, South Synopsis
Africa. Ferrochrome (FeCr) is a vital ingredient in stainless steel production and is commonly produced
3Department of Chemical Engineering, by smelting chromite ores in submerged arc furnaces. Silicon (Si) is a componrnt of the FeCr alloy
Stellenbosch University, South Africa. from the smelting process. Being both a contaminant and an indicator of the state of the process,
its content needs to be kept within a narrow range. The complex chemistry of the smelting process
and interactions between various factors make Si prediction by fundamental models infeasible. A
Correspondence to: data-driven approach offers an alternative by formulating the model based on historical data. This
A.V. Cherkaev paper presents a systematic development of a data-driven model for predicting Si content. The
model includes dimensionality reduction, regularized linear regression, and a boosting method
Email: to reduce the variability of the linear model residuals. It shows a good performance on testing
[Link]@[Link] data (R2 = 0.63). The most significant predictors, as determined by linear model analysis and
permutation testing, are previous Si content, carbon and titanium in the alloy, calcium oxide in
the slag, resistance between electrodes, and electrode slips. Further analysis using thermodynamic
Dates: data and models, links these predictors to electrode control and slag chemistry. This analysis lays
Received: 31 Aug. 2022 the foundation for implementing Si content control on a ferrochrome smelter.
Revised: 11 Dec. 2022
Accepted: 13 Dec. 2023 Keywords
Published: February 2024
ferrochrome, silicon content, machine learning, principal component analysis, gradient boosting,
mutual information.
How to cite:
Cherkaev, A.V. Erwee, M., Reynolds, Q.G.
and Swanepoel, S. 2024
Prediction of silicon content of alloy in
ferrochrome smelting using data-driven Introduction
models.
Journal of the Southern African Institute Ferrochrome (FeCr) is an alloy commonly used as raw material in the production of stainless steel and
of Mining and Metallurgy, vol. 124, special steels requiring corrosion and creep resistance. The largest deposits of chromite ore occur in South
no. 2. pp. 67–74 Africa (Haldar, 2020). Most of the ferrochrome is produced by means of fluxed smelting of chromite ore
using quartz and limestone in a submerged arc furnace (SAF). Typically, FeCr produced in this process
DOI ID: contains about 50−53% Cr, 4−6% Si, and 6-8% C with the balance being Fe (Gasik, 2013).
[Link] The reduction of chromite with carbon in SAFs is quite complex, and has been documented in detail by
9717/2297/2024 several authors (Xiao, Yang, and Holappa, 2006; Hayes, 2004; Ringdalen and Eilertsen, 2001; Wedepohl and
Barcza, 1983), but the overall reactions can be simplitied as follows:
ORCID:
A.V. Cherkaev [1]
[Link]
[2]
[3]
Typically, slag is tapped at temperatures above 1700°C, with an average in-furnace temperature
in the order of 1800°C. At these temperatures, the undesirable reduction of SiO2 (Equation [3]) is
thermodynamically favourable and directly in competition with the first two chromite reduction reactions.
The alloy thus becomes contaminated with Si. Aside from the quality issue, every 1% increase in Si content
adds some 45 kWh per ton alloy to the specific energy requirement for smelting (Ringdalen, Rocha, and
Figueiredo, 2015). In contrast, a low Si content in the alloy indicates a lack of active carbon for reduction
reactions (Urquhart, 1972). It is for these reasons that understanding how to control Si in SAF smelting of
chromite is important.
A target Si content in the alloy of no more than 4.5% is desirable as per general customer specifications.
The exact mechanisms that can lead to variation of Si in the alloy are quite complex, and beyond the scope
of this paper. The reader is referred to the literature for more information on the topic (see Ranganathan et
The Journal of the Southern African Institute of Mining and Metallurgy VOLUME 124 FEBRUARY 2024 67
Prediction of silicon content of alloy in ferrochrome smelting using data-driven models
al., 2005; Ranganathan and Godiwalla, 2011; Hockaday and Bisaka, Anatomy of the process monitoring data
2010; Geldenhuys, 2013; Erwee, Swanepoel, and Reynolds, 2021). A Process monitoring data can be either numerical (temperature,
simplistic summary of these effects in SAF smelting of chromite is flow rate, pressure, etc.) or categorical (e.g., type of feed). The latter,
given in Figure 1. It is important to note that the diagram does not however, is rarer compared to numerical data. One common feature
highlight all the interactions between the different causes; these are of all process monitoring data is that each variable is a time series,
alluded to in the literature cited above. Due to the complexity of the i.e. the order of observations matters. Therefore, ML models should
process, machine learning methods can offer not only insight into not only consider present values of the variables, but also (in some
the relative importance of the drivers for Si content of the alloy, but form) their past values.
also a possible means of control by way of supporting decisions-
made by operators. ML model framework
Machine learning methods for process monitoring The following are common steps in a (supervised) ML model
application:
Machine learning (ML) methods have been applied for process
monitoring tasks in various industries for over a decade (McCoy ➤ Data cleaning: removal of noise, imputing missing values, pre-
and Auret, 2019) to support decision-making by operators and, selection of the features (usually based on domain knowledge)
where possible, to automate it. At a more technical level, ML ➤ Splitting of the data into training and testing data-sets
models are used for such tasks as data cleaning, missing data ➤ Transformation of variable values (standardization)
imputation, and noise reduction (unsupervised methods that ➤ Feature engineering: addition or replacement of variables
include dimensionality reduction and clustering), and soft sensing using statistical or fundamental principles
and forecasting (supervised methods that include regression ➤ Dimensionality reduction (data simplification): densification
and classification). Unsupervised methods are characterized by of information content in the data
the absence of the output, or response, variable i.e., the target ➤ Application of a regression or classification model.
is unknown. Their aim is to simplify and condense the data. In Depending on the task and context, some intermediate steps
contrast, supervised methods aim to infer or predict a particular may be skipped. Below is a short description of what each of these
response, e.g., temperature in a reactor, liquid level in a vessel, steps may involve. Step 6 is discussed in more details in the next
product composition, etc. subsection.
The rest of this subsection describes the process of ML model Obvious outliers, resulting from malfunctioning sensors
application to industrial data, starting with the terminology as this
or clerical errors in the case of manual data input, can be and
is quite specific to the ML field.
should be removed from the data as they can severely affect model
Terminology performance. However, identification of such outliers for a complex
The data-set used for ML models is a table of entries consisting nonstationary process is a challenge and no general reliable
of M rows and N columns. Each column corresponds to a feature methods exist. Missing values are a common occurrence in process
being measured. For example, hearth temperature, current through monitoring data, especially in historical data where a particular
electrode, type of reductant, etc. A column is referred as a variable. variable was not measured from the beginning of the recording
A variable that is aimed to be inferred by an ML model is referred period. Many ML methods cannot deal with missing values. Thus,
as a target, response, or output variable. Other variables are called it is desirable to replace them with some sensible numeric values.
predictors or features. The number of columns N determines the Although several methods of imputation are available (use overall
dimensionality of the data (target variables are usually excluded mean, last value carried forward, use some regression model to
from it). Each row of the table corresponds to an individual estimate missing values), each method comes with serious trade-
observation and, thus, the term observation is used to denote a row offs and, thus, imputation is often skipped. There is a danger of
(James et al., 2017). information leakage during data cleaning. To avoid it, this step is
Using a probabilistic view, each column is modelled as a sometimes performed after step 2.
random variable (RV). The values in the column provide concrete The ML model is fitted to the data. However, the same data
realizations of this variable. Hence, it is possible to use usual cannot be used to assess model predictive performance. Thus, the
characteristics of RVs such as distribution functions, means, data is split into a training set and a testing set. The former is used
variances, etc. Furthermore, from this viewpoint ML is the same as in learning algorithms, whereas the model predictive powers are
statistical inference and aims at inferring probabilistic properties of examined using the latter. In general, the observations are sampled
the RVs. randomly without replacement to either training or testing parts.
Figure 1—Fishbone diagram illustrating factors leading to excess Si in FeCr alloy produced via the SAF smelting route
68 FEBRUARY 2024 VOLUME 124 The Journal of the Southern African Institute of Mining and Metallurgy
Prediction of silicon content of alloy in ferrochrome smelting using data-driven models
However, in the case of time series data this is not possible since in the present paper) or to classify a target into a particular group
observations are not independent. In that case, observations are (e.g., alarm or no-alarm condition). A commonly used term for
split at some time point T, with all the observations preceding the former is regression, and for the latter, classification. Since the
T reporting to the training set and the rest to the testing set. As model developed in this paper is a regression model, the rest of this
already mentioned, it is advisable to perform this split as early as subsection is focused on regression.
possible to avoid information leakage. Linear regression models are among the oldest in ML. Despite
Many ML methods involve linear combination of features. If the their simplicity, they can still produce satisfactory inferences or
scales of values of variables differ, which is the case when variables forecasts for processes that are tightly controlled. To improve model
correspond, for example, to temperatures, flow rates, or weights, it generalization capabilities, its flexibility is sometimes restricted by
can mislead model training and skew the results. A few methods the method called regularization. LASSO (least absolute shrinkage
of scaling are available. Standardization refers to shifting the values operator) linear regression is a regularization method that restricts
such that their mean is zero and scaling to achieve a standard the magnitude of linear regression coefficients in such a way that the
deviation of unity. Normalization linearly maps the original range of coefficients for the predictors that are only weakly correlated with
values to a [0,1] or [–1,1] interval. Normalization is useful if values the target are set to zero (James et al., 2017). Thus, it can be used
have hard bottom and top values (for example, liquid levels in a not only to improve model performance, but also to assess variable
tank). importance.
Step 4 augments the data with new features. There can be a Despite their robustness, linear methods can miss noninear
physical basis for additional variables. For example, pump gain, effects and variable interactions. Tree-based methods with bagging
defined as a ratio of flow rate to pressure produced, can be added (random forest) or boosting (gradient boosting) are among the
to the data. Statistically based variables are often used to good most competitive methods for nonlinear inference and forecasting
effect. For example, high-frequency data can be replaced by lower- (Markidakis et al., 2022). A tree-based method uses training data
frequency observations of the mean, variance, or higher order to recursively partition the input space to regions in which the
moments over a time window. Additionally, time-shifted values target variable is approximately constant. An individual tree-
of existing variables can be incorporated as additional features to based predictor is very sensitive to slight changes in the input data
capture the process dynamics (time embedding; Marwan et al., and tends to over-fit the model (James et al., 2017). Bagging and
2007). More precisely, for a time series variable Xt an additional set boosting are two methods that help to overcome this weakness
of variables [Xt-h, Xt-2h ,…, Xt-kh] is added in such a way that each by constructing multiple trees on different portions of data and
(scalar) observation x(t) of Xt is replaced by a vector (x(t), x(t-h), then aggregating their results. They are known as ensemble
x(t-2h) ,…, x(t-kh)) for some time shift step h and embedding methods. Such methods combine multiple weak learners to form
dimensionality k. Such augmentation then naturally leads to auto- a single strong learner. Although both bagging and boosting tend
regressive (AR) type models (Chatfield, 1996). If the embedding is to perform similarly, recent studies are showing that boosting
used to capture the process dynamics, it results in each observation outperforming bagging on a variety of problems (Makridakis et
becoming independent from others. Thus, if it is done before the al., 2022). AdaBoost, XGBoost, and LigthGBM are commonly
training-testing data split, the split can be performed by randomly used, freely available implementations of the boosting algorithms.
Among them, recently-developed LightGBM provides much shorter
sampling individual observations.
training times (Luo and Chen, 2020).
Dimensionality reduction (step 5) is a common preprocessing
task in process monitoring as it helps to remove redundant variables Model analysis: mutual information and variable
from the data-set. Principal component analysis (PCA) is a standard importance
technique for this due to its simplicity and robustness. PCA finds
Typical regression model performance metrics include root mean
the set of latent variables, also known as principal components,
square error (RMSE)
that are linear combinations of the original variables in such a way
that (i) each latent variable is uncorrelated from all others, and (ii)
principal components are sorted in the descending order of their
variability. Thus, it becomes possible, by selecting only a few first
principal components, to account for most of the variability of the
predictors. Other techniques for dimensionality reduction include
mean absolute error (MAE)
independent component analysis (ICA) (Hyvärinen and Oja, 2000),
linear discriminant analysis (LDA), empirical mode decomposition
(EMD) (Boudraa and Cexus, 2007), and self-organizing maps
(SOMs), to name a few. These methods are usually applied only
if PCA fails to produce desirable outcomes or if there is a more and coefficient of determination R2
specific objective, such as identification and separation of different
sources in the signal (ICA) or identification of the clusters (LDA)
(James et al., 2017). The main result of step 5 is densification of
information content in the data-set, which helps to reduce the Here, yi is the observed target value, y− is the average target
complexity of a subsequent regression or classification model (step value, and y^i is the inferred (predicted) target value. These metrics
6). give an overall summary of how good the model inference is.
However, there are some details that they cannot capture. For
Linear and nonlinear regression models example, although they can tell whether model predictions are of
In supervised learning there are two main objectives: either to ‘good’ or ‘bad’ quality, they cannot tell if the model can be improved
predict a particular value of the target (such as Si content in alloy as (using current data).
The Journal of the Southern African Institute of Mining and Metallurgy VOLUME 124 FEBRUARY 2024 69
Prediction of silicon content of alloy in ferrochrome smelting using data-driven models
Mutual information (MI) is a useful tool to assess the The results of the analysis of the chemical compositions of
information content of predictors with respect to the response. MI tapped slag and alloy are recorded in tap data. Tap data is considered
between two random variables X and Y, I[X,Y], is defined as follows: ‘medium frequency’ in this study as it is recorded on average every
3.2 hours. Since the analysis requires some time to perform, this
data is not recorded online. The tap data contains 70 variables and
11 648 observations.
Recipeand KPI data (recipes data for short) is recorded daily
where integration is performed over the range of values of X and and summarizes, among other indicators, the amounts and sources
Y and pX, pY, and pXY are probability density functions of X, Y, and of different raw materials (reductant, flux, and ore) used. Since
their joint distribution respectively (Paninski, 2003). If variables X there are many sources and categories of, for example, reductant,
and Y are independent, it follows that I[X,Y] 0, since pXY = pX pY recipe data is sparse: 57% of all entries are zero. The recipe data-set
and the logarithm argument becomes unity. MI reaches a maximum contains 156 variables and 1556 observations.
when there is a purely functional relationship between the variables:
by observing X one can precisely deduce Y. MI can be used to pre- Data-set preparation
select the predictors or examine how important is the contribution
of an individual predictor. Embedding
To facilitate the comparison between MI that corresponds to Given the long processing times of a typical SAF, it is expected
different predictors, it is useful to rescale MI and introduce relative that not only most recently observed feature values, but also
MI, RMI, as follows (Cherkaev et al., 2022): past values of features and the target, will influence future target
value. To account for this influence, time embedding (see ML
model framework subsection for more details) on the data-set
was performed. It requires choosing the time shift step (delay
interval) h and dimensionality k of the embedding. Although there
which ranges from 0 to 1 (or 0% to 100%). Thus, if RMIs between are statistical methods to determine these parameters, they rely
the response and all the predictors are low, it is not expected of any
model to provide a useful inference of the response (‘garbage in—
garbage out’).
Study objectives
This study aims to predict Si content in the tapped FeCr alloy
using the data available up to the point when the tap starts. The
methodology is similar to that of (Jonsson and Ingason (1998), but
it extends the approach in the following ways:
➤ Si content is predicted per tap, not per day
➤ A wide range of predictors is used. Regularization methods
are used to identify important predictors
➤ A nonlinear model is used to improve linear model results
➤ Variable importance is assessed.
The rest of the paper is structured as follows. An overview of
the data used for the study is followed by a description of the steps
taken for data preparation (steps 1–4 of the ML model framework).
Dimensionality reduction and regression model development are
sthen shown, followed by an assessment of the model performance
and variable importance. Analysis of residuals assesses the
possibility of improving the model. Finally, a general discussion and
conclusions are presented.
Process monitoring data in FeCr smelting
Furnace process monitoring data comprises three distinct aspects: Figure 2—Pair plot of the selection of the variables
regular process monitoring data (named here tag data), tap data,
and recipe and KPI (key performance indicator) data.
The outputs of temperature and electrical signal sensors are
recorded at high frequency to tag data. Here ‘high frequency’ means
that data is recorded most often. In many operations such data is
recorded manually on an hourly basis. In this study, tag data was
recorded automatically and sampled every 2 minutes. The recorded
data includes hearth and shell temperatures, electrode currents,
electrode holder positions, and power attributed to each electrode.
Due to malfunction of some of the sensors, the data-set contains
a considerable number of missing entries. Overall, it contains 130
variables and 2 412 000 observations. Figure 3—Time series plot of Si and C content in metal
70 FEBRUARY 2024 VOLUME 124 The Journal of the Southern African Institute of Mining and Metallurgy
Prediction of silicon content of alloy in ferrochrome smelting using data-driven models
heavily on stationarity of the time series. Technical data in the Standardization
processing industry is rarely stationary in the statistical sense, and The training data-set was used to calculate the mean and variance
these methods often fail. A more reliable way to determine these for each variable. These values were used to standardize all data-sets
parameters is to use domain-specific knowledge. For example, it – training, testing, and validation.
cannot be expected to see the effects of recipe changes past two
days. As for the tap data, a study on ferromanganese smelting Model development
(Cherkaev et al., 2022) showed that it is not necessary to choose the This section presents the development of the data-driven model to
dimensionality k > 2. predict Si content. All the development is done using the training
Tag data is not embedded in this study since tag data represents data-set.
immediate furnace conditions that have only a short-term effect.
Tap data is embedded with n = 2, i.e., two previous taps are Dimensionality reduction
considered with the current tap entry. Similarly, recipe data is The high dimensionality of the data is a serious issue as it causes
embedded with dimensionality n = 2. the variable-space to be only sparsely populated by observations.
One notable consequence of embedding is the considerable Furthermore, it is likely that the data contains many correlated
increase in the number of variables. Both, tap data and recipe data variables. PCA was performed on the training set. Since PCA
increased the number of variables three times for n = 2. cannot process missing values directly, observations that contain
Synchronization them were removed from the data-set. The dimensionality reduction
was performed by retaining the first 124 components that account
Working with three separate data-sets is not convenient. Thus, for 90% of the total variance.
the data-sets were combined to form a single set. Tap data was
used as the base. Since there are many observations in tag data Linear model
corresponding to a single entry in tap data, i.e., there are many Linear models have two attractive properties: they are cheap to train
observations from the previous tap start to the current tap start, and easy to interpret. Thus, if there is a linear relationship between
these were replaced by a mean value. Furthermore, to account predictors and the response (Si content), it is advisable to extract it
for tag data variability, an additional variable was introduced: the using a specialized linear model instead of relying on more general
standard deviation of the original variable between the taps. To nonparametric models.
summarize, each tag variable X with values xn, xn+1 ,…, xn+k between Similarly to PCA, linear models such as ordinary linear least
the tap opening times is replaced by two variables: squares (OLS), regularized OLS, or partial least squares (PLS)
cannot handle missing values directly. The same approach as for
PCA was followed: only the observations without missing values
were used for training. LASSO linear regression was chosen as a
and linear model since it produces a sparse model (reduces the number
of predictors), which helps model interpretation. The shrinkage
parameter was automatically tuned during the learning procedure.
The LASSO model was able to reduce the number of predictors
from 124 to 84.
The LASSO model was able to reduce the variance of Si content
each containing a single observation per tap. Recipe data was on the training data-set from 1 (in standardized form) to 0.49.
extended with zero-hold (last value carried forward) to provide one The distribution of residuals for the training data-set is shown in
entry per tap. Figure 4. There is still a significant variability of residuals, which
Finally, current day recipe and current tap entries were excluded it is desirable to reduce. For this a nonlinear model needs to be
from the set of features since they cannot act as predictors for the employed.
current tap: they are available either at the end of the day or after the
current tap is complete. Gradient boosting
As a result of synchronization, dropping the variables that It is assumed that the model is additive:
contain too many missing values and dropping the rows with
missing tap data, the combined data-set contains 745 variables and
11 511 observations.
Training, testing, and validation data
The combined data-set was split into training, testing, and final
validation sets. Thanks to time embedding performed on tap and
recipe data, each entry in the data-set is independent of previous
entries. Thus, training and testing samples can be drawn randomly
from the data. For the final assessment of the model performance it
is, however, more instructive to draw a continuous sample.
First, 200 contiguous observations were drawn from the middle
of the data-set for the final validation data; 70% of the rest of the
observations were randomly chosen for training, and the remainder
were assigned for testing. This resulted in training and testing data Figure 4—Histograms of Si content, linear model residuals, and final model
having 8057 and 3254 observations respectively. residuals for the training data
The Journal of the Southern African Institute of Mining and Metallurgy VOLUME 124 FEBRUARY 2024 71
Prediction of silicon content of alloy in ferrochrome smelting using data-driven models
where residuals contain variability that was not explained by a linear
model. These residuals are modelled using a gradient boosting (GB)
method as implemented by the LightGBM package. GB involves
a number of hyperparameters such as depth of individual trees,
learning rate, and regularization parameters. A grid search was
employed with 5-fold cross-validation to fine-tune these parameters.
The final model was trained on the full training data-set with the
fine-tuned hyperparameters.
Model performance
Figure 5—Histograms of Si content, linear model residuals, and full model
Model performance on training data residuals for testing data
Although it is not possible to judge model performance when
the model is applied to training data, it provides the base for
comparison when applied to the testing data. Furthermore, it can
give an idea of how effective model training was.
The coefficient of determination R2, MAE, and RMSE for both
the linear model and the combined model are shown in Table I.
Figure 4 compares the distributions of Si content, the linear model
residuals, and the combined model residuals. It is evident that each
model provides a considerable reduction in residuals variability.
This is especially true for the addition of the GB model.
Model performance on testing data
Performance on the testing data is shown in Table II and Figure 5.
The performance indicators of the linear model on the testing data Figure 6—Comparison of the model predictions and true Si content on the
contiguous validation data-set
are similar to those on the training data. In contrast, the combined
model shows much less variability reduction compared to the identify the most important predictors of the model output. Linear
training data. Nonetheless, there is still considerable reduction in models, including regularized models such as LASSO, offer a direct
variability by both linear and nonlinear models. way for this by inspecting the model coefficients. GB models offer
Validation on contiguous data variable importance ranking as a part of the learning algorithm.
However, for the current study this method is suboptimal since a
The result of the model applied to the contiguous data is shown in
GB model is used to improve the linear model results. A more direct
Figure 6. It is evident that although some variability in measured
approach for ranking variable importance is to use a permutation
Si content is missed, model predictions provide an improved and
method, by which the observations of each variable are shuffled and
reliable estimate compared to mean value (zero).
the reduction in model accuracy is assessed.
Importance of variables Features selected by the LASSO linear regression model
For the purpose of Si content control, it is not enough to only Since the LASSO model was applied to the result of PCA
predict it, but it is necessary to identify factors that affect it. While transformation, they need to be considered together. PCA
it is possible to point out several physical factors that control Si transformation can be represented as follows:
content, it is difficult to identify a particular factor that drives it
up or down. Machine learning models provide a few techniques to
where X M×N is the data matrix with M observations of N
Table I variables, R N×N is the full (orthogonal) rotation matrix, and
Model accuracy on training data P M×N is the matrix of principal components. The LASSO
model has the form
Metric Linear model Combined linear and GB model
R2 0.50 0.77
MAE 0.38 0.27 where Y is the response (Si content), Pj is the j-th column of P, βj
RMSE 0.70 0.47
are non-zero coefficients of the LASSO model, and the ellipsis is
used to show that other components were omitted. Combining this
Table II equation with PCA transformation gives:
Model accuracy on testing data
Metric Linear model Combined linear and GB model
where Ri is the i-th column of matrix R. Since all the variables in
R2 0.53 0.63 matrix X have zero mean and variance unity (after standardization),
MAE 0.38 0.32 their contribution can be assessed by the magnitude of the entries of
RMSE 0.64 0.57
the vector
72 FEBRUARY 2024 VOLUME 124 The Journal of the Southern African Institute of Mining and Metallurgy
Prediction of silicon content of alloy in ferrochrome smelting using data-driven models
Selected features and thermodynamics
The features from the model, specifically the carbon and titanium
The first four largest entries of vector C correspond to Si content contents of the alloy in the previous tap, the CaO content and
in the last two taps and average Si content in the last two days. This basicity of the slag, and electrode parameters reflect what
shows the strong autoregressive nature of the model. Other strong fundamentally drives many reduction reactions in pyrometallurgical
predictors include carbon and titanium content in metal in the processes, i.e. the relative effects of chemistry and temperature.
previous tap, CaO content in the slag, and slag basicity. Consider the main reaction for SiO2 reduction in the furnace,
Variables importance of the full model repeated here for convenience:
The top importance predictors that were identified by the
permutation importance method include past Si content values, [4]
carbon content in the previous tap (the second most important),
electrode resistance, and electrode slip. The equilibrium constant for the reaction is:
Analysis of residuals [5]
Normality test
The Shapiro-Wilk normality test indicates that the combined model The effect of temperature on the equilibrium constant for
residuals are not Gaussian (p-value is 1.25×10–25). Therefore, it the reaction is given in Figure 7a. Although a simplistic view of a
is likely that there are a small number of factors that can further complex process, the temperature at which SiO2 reduction with
reduce their variability and, thus, improve the model. carbon occurs has been shown to be almost overriding in SAF
smelting of chromite (Haldar, 2020; Hayes, 2004) and largely due to
Mutual information improper heat input due to poor electrode control (James, Witten,
Mutual information analysis was performed on the testing data-set Hastie, and Tibshirani, 2017) , i.e. if the resistance of the burden
to determine if there are predictors in the current data-set that can is such that the electrode tips sit high up in the furnace, a greater
have a functional relationship with residuals. Predictors with more amount of energy is spent on a small volume of material, readily
than 50% of missing values in the testing data-set were excluded increasing the Si content of the alloy. Once Si is produced in a SAF
from this analysis. process, the alloy droplets trickle to the bottom of the furnace. Very
Compared to a fully deterministic model (i.e., residuals vs. little can be done to refine the metal further, unlike, for example in
residuals) taken as 100% MI, the largest MI between the residuals open-bath chromite smelting processes (Hyvärinen, and Oja, 2000),
and predictors is 0.7%, corresponding to the active power standard where the alloy can still be refined by addition of highly basic oxides
deviation between the last and the current taps. For comparison, the such as CaO.
top relative MI between the predictors and Si content is 7.5%. Low This said, there is still a sound basis for the addition of basic
values of MI between the predictors for model residuals indicate oxides (e.g. CaO) to compensate for high Si contents of the alloy – if
that it is unlikely that the current model can be improved using the a mini-melt of high basicity slag forms close to the reaction sites for
same predictors. SiO2-C reaction, the activity of SiO2 in the melt can be suppressed
to some degree by having a more basic slag, as shown in Figure 7b,
Discussion limiting the extent to which the SiO2-C reaction is driven.
Reducible oxides in the slag, e.g. TiOx and VOx, naturally
Effect of recipe compete with CrOx and SiO2 in the slag for carbon. It is therefore
One curious outcome of the features importance analysis is the not surprising that some of the minor elements in the alloy such
absence of any recipe-related variables. The most likely explanation as Ti come out as predictors for Si. Interaction between the alloy
for this is that the recipe is correctly designed for the process. components can be described thermodynamically; however, the
Indeed, recipe adjustments are made to control, among other complexity and nonequilibrium nature of the SAF smelting process
parameters, Si content in the metal. Therefore, for the learning make it more difficult to do so. It is for this reason that a machine
procedure, changes in the recipe appear to have no effect on Si learning approach becomes invaluable. The effect of temperature
content. The best way to pick up the recipe effect would be to (driven by electrical energy input from the electrodes) and
randomly change the recipe, or at least to forgo Si content control in chemistry effects are neatly combined into a practical model that
the recipe calculation. This, however, is impractical. can be used on the plant.
Figure 7—(a) Effect of temperature on the equilibrium constant for the SiO2-C reaction and (b) effect of slag basicity on the activity of SiO2 in a MgO-Al2O3-CaO-SiO2
slag in SAF FeCr smelting at 1700°C. B3 = (%CaO+%MgO)/(%SiO2). Calculations done in FactSage 8.0, using the FTOxid database
The Journal of the Southern African Institute of Mining and Metallurgy VOLUME 124 FEBRUARY 2024 73
Prediction of silicon content of alloy in ferrochrome smelting using data-driven models
Conclusions Haldar, S.K. 2020. Minerals and rocks. Introduction to Mineralogy and Petrology.
Elsevier. pp. 1–51.
This study aimed to develop a data-driven Si content model and
Hayes, P.C. 2004. Aspects of SAF smelting of ferrochrome. INFACON X:
identify important drivers of Si content in FeCr smelting. The
‘Transformation through Technology’. Proceedings of the Tenth International
model, based on time-embedding, PCA, LASSO regression, and Ferroalloys Congress, Cape Town, South Africa, 1–4 February 2004. https://
gradient boosting regression tree methods showed good predictive [Link]/InfaconX/[Link]
capabilities. The Si, C, and Ti contents in the metal and CaO Hockaday, S.A.C. and Bisaka, K. 2010. Some aspects of the production of
content in the slag from the previous taps are among the strongest ferrochrome alloys in pilot DC arc furnaces at Mintek. INFACON XII.
predictors of the Si content at he next tap. The importance of these Proceedingas of the 12th International Ferroalloys Congress, Helsinki, Finland.
predictors can be explained by the ‘inertia’ of a SAF (previous Si Outotek Oyj. pp. 367–376.
content) and chemical properties of the ore, reductant, and flux Hyvärinen, A. and Oja, E. 2000. Independent component analysis: algorithms and
mixture. The absence of recipe-related variables as important applications. Neural Networks, vol. 13. pp. 411–430.
features is probably due to the recipe being correctly designed. James, G., Witten, D., Hastie, T., and Tibshirani, R. 2017. An Introduction to
Analysis of residuals shows that it is unlikely that the current model Statistical Learning with Applications in R. Springer.
can be improved any further (given the same data-set). Overall, Jonsson, G.R. and Ingason, H.T. 1998. On the control of silicon content in
this model can give an early warning sign to the operator if the Si ferrosilicon. INFACON VIII. Proceedings of the 8th International Ferroalloys
content is moving away from the target. Congress, Beijing, China, 7-10 June 1998. China Science & Technology Press.
pp. 95–98.
Acknowledgements Luo, S. and Chen, T. 2020. Two derivative algorithms of gradient boosting decision
The financial assistance of the National Research Foundation (NRF) tree for silicon content in blast furnace system prediction. IEEE Access, vol. 8.
pp. 196112–196122.
toward this research is hereby acknowledged. Opinions expressed
and conclusions arrived at are those of the author and are not Makridakis, S., Spiliotis, E., Assimakopoulos, V., Chen, Z., Gaba, A., Tsetlin,
necessarily to be attributed to the NRF. I., and Winkler, R.L. 2022. The M5 uncertainty competition: results, findings
and conclusions. International Journal of Forecasting, vol. 38, no. 4. pp. 1365-
This paper is published by permission of Mintek. 1385.
Conflict of interest Marwan, N., Romano, M.C., Thiel, M., and Kurths, J. 2007. Recurrence plots for
the analysis of complex systems. Physics Reports, vol. 438. pp. 237–329.
On behalf of all authors, the corresponding author states that there
McCoy, J.T. and Auret, L. 2019. Machine learning applications in minerals
is no conflict of interest. processing: a review. Minerals Engineering, vol. 132. pp. 95–109.
CRediT author statement Paninski, L. 2003. Estimation of entropy and mutual information. Neural
Computation, vol. 15. pp. 1191–1253.
AVC: Methodology – Machine learning, Investigation, Formal
Ranganathan, S. and Godiwalla, K.M. 2011. Influence of process parameters
analysis, Software, Writing, Visualisation, Project administration; on reduction contours during production of ferrochromium in submerged arc
ME: Methodology – Thermodynamics, Investigation, Writing, furnace. Canadian Metallurgical Quarterly, vol. 50. pp. 37–44.
Visualisation; QGR: Supervision, Funding; SS: Conceptualisation,
Ranganathan, S., Mishra, S.N., Mishra, R., and Singh, B.K. 2005. Control
Supervision. of silicon in high carbon ferrochromium produced in submerged arc
furnace through redistribution of quartzite in the charge bed. Ironmaking &
References Steelmaking, vol. 32. pp. 177–184.
Boudraa, A-O. and Cexus, J-C. 2007. EMD-based signal filtering. IEEE Ringdalen, E. and Eilertsen, J. 2001. Excavation of a 54 MVA HC-ferrochromium
Transactions on Instrumentation and Measurement, vol. 56. pp. 2196–2202. furnace. INFACON IX. Proceedings of the Ninth International Ferroalloys
Chatfield, C. 1996. The Analysis of Time Series. An Introduction. 5th edn. Congress. Quebec City, Canada, 3-6 June 2001. The Ferroalloys Association,
Chapman & Hall/CRC. Washinhton, DC. pp. 166–173. [Link]
InfaconIX/[Link]
Cherkaev, A.V., Rampyapedi, K., Reynolds, Q.G., and Steenkamp, J.D. 2022.
Tapped alloy mass prediction using data-driven models with an application to Ringdalen, E., Rocha, M., and Figueiredo, P. 2015. Energy consumption during
silicomanganese production. Furnace Tapping 2022. Springer. pp. 131–144. HCFeCr-production at Ferbasa. INFACON XIV. Proceedings of the Fourteenth
International Ferroalloys Congress, Kiev, Ukraine, 31 May-4 June 2015.
Erwee, M., Swanepoel, S., and Reynolds, Q. 2021. The importance of controlling
pp. 668–675. [Link]
the chemistry of pre-oxidized chromite pellets for submerged arc furnace FeCr
smelting: a study on furnace Si control. Proceedings of the 16th International Urquhart, R. 1972. The production of high-carbon ferrochromium in a
Ferro-Alloys Congress (INFACON XVI) ,Trondheim, Norway, 27-29 September submerged-arc furnace. Minerals Science and Engineering, vol. 4. pp. 48–65.
2021. [Link]
Wedepohl, A. and Barcza, N.A. 1983. The 'dig-out' of a ferrochromium furnace.
Gasik, M.M. 2013. Introduction. Handbook of Ferroalloys. Elsevier. pp. 3–7. Special Publication no, 7. Geological Society of South Africa, Johannesburg.
Geldenhuys, I.J. 2013. Aspects of DC chromite smelting at Mintek — An Xiao, Y., Yang, Y., and Holappa, L. Tracking chromium behaviour in submerged
overview. Proceedings of the Thirteenth International Ferroalloys Congress, arc furnace for ferrochrome production. Proeceedings of the Sohn International
Almaty, Kazakhstan, 9-13 June 2013. [Link] Symposium. Volume 7- Industrial Practice. Kongoli, F. and Reddy, R.G. (eds).
InfaconXIII/[Link] TMS, Warrendale, PA. pp. 417-433. u
74 FEBRUARY 2024 VOLUME 124 The Journal of the Southern African Institute of Mining and Metallurgy