Minerals Engineering 162 (2021) 106718

Contents lists available at ScienceDirect

Minerals Engineering
journal homepage: www.elsevier.com/locate/mineng

Modelling for froth flotation control: A review

Paulina Quintanilla *, Stephen J. Neethling, Pablo R. Brito-Parada
Department of Earth Science and Engineering, Royal School of Mines, Imperial College London, South Kensington Campus, London SW7 2AZ, United Kingdom

A R T I C L E I N F O A B S T R A C T

Keywords: Flotation is a conceptually simple operation; however, as a multiphase process with inherent instability, it ex­
Froth flotation hibits complex dynamics. One of the most efficient ways to increase flotation performance is by implementing
Flotation control advanced controllers, such as Model Predictive Control (MPC). This type of controller is very dependent on the
Flotation modelling
model that represents the dynamics of the process. Although model development is one of the most crucial parts
Model predictive control
in MPC, flotation models have been mainly developed for simulation purposes (i.e. analysis and design) rather
than control purposes. This paper presents a critical literature review on modelling for froth flotation control.
Models reviewed have been sub-classified as empirical, phenomenological and hybrid according to their char­
acteristics. In particular, it is highlighted that models have so far primarily focused on the pulp phase, with the
froth phase often neglected; when the froth phase is included, kinetics models such as those used for the pulp
phase, are commonly used to represent it. Froth physics are, however, dominated by processes such as coales­
cence, liquid motion and solids motion, which have been previously modelled through complex, steady-state
models used for simulation purposes, rather than control purposes. There remains a need to develop appro­
priate models for the froth phase and more complex models for the pulp phase that can be used as part of MPC
strategies. The challenges associated with the development of such models are discussed, with the aim of
providing a pathway towards better controlled froth flotation circuits.

1. Introduction and the velocity and stability of the froth; and the mineral concentration
in the feed, concentrate and the tailings (Laurila et al., 2002). In terms of
Froth flotation is the largest tonnage separation in mineral process­ process control, these variables are classified as manipulated, distur­
ing by which valuable mineral is separated from waste rock. Advances in bance, controlled, and internal state variables.
control and optimisation of the froth flotation process are of great Manipulated variables are defined as those that can be modified to
relevance since even very small increases in recovery lead to large change the internal states of the model; disturbance variables are those
economic benefits (Ferreira and Loveday, 2000; Maldonado et al., that cannot be modified or controlled and, only in some cases, can be
2007). However, the implementation of advanced control and optimi­ measured or estimated; controlled variables are the objective of the
sation strategies is not always completely successful in flotation. This is control; and state variables are internal variables that define the models
because flotation performance is affected by a great number of variables (Sbarbaro and del Villar, 2010). Particularly for flotation, control vari­
that interact with each other (Arbiter and Harris, 1962; Laurila et al., ables are usually classified as follows (Hodouin, 2011; Jovanović et al.,
2002), while unmeasurable disturbances in the process further compli­ 2015; Shean and Cilliers, 2011) (as shown in Fig. 1):
cate the implementation of efficient strategies (Cubillos and Lima,
1997). • Inputs (or independent variables):
Froth flotation is affected by several variables, such as the flowrates – Manipulated variables: air flowrate, reagent addition rate,
into the flotation cell, i.e. feed, air and froth wash water, as well as the tailings flowrate, feed mass florwate, wash water flowrate
addition rate of the various chemical reagents; the slurry density and (specially in columns), frother addition rate.
solids content as well as the slurry level in the tank; electrochemical – Disturbances: particle size distribution, solids content, head
potentials such as pH, Eh and conductivity; ore mineralogy and the size, grade, volumetric flowrate, particle properties (size distribution,
distribution and liberation of the particles; the bubble size distribution

* Corresponding author.
E-mail address: [email protected] (P. Quintanilla).

https://doi.org/10.1016/j.mineng.2020.106718
Received 21 April 2020; Received in revised form 28 September 2020; Accepted 28 November 2020
Available online 29 December 2020
0892-6875/© 2020 Elsevier Ltd. All rights reserved.
P. Quintanilla et al. Minerals Engineering 162 (2021) 106718

Fig. 1. Classification of flotation variables. (Adapted from Lynch et al. (1981), Hodouin (2011), Jovanović et al. (2015)).

mineralogy, shape, degree of liberation), froth properties, elec­ review is not focused on instrumentation itself, and the reader is referred
trical potentials in the pulp (i.e. pH and Eh). to Laurila et al. (2002) and Shean and Cilliers (2011) for a full
• Outputs (or dependent variables): description of instrumentation in froth flotation plants.
– Controlled variables: pulp level, froth depth, grade, gas hold-up. The most conventional controller is the so-called Proportional-Inte­
– Internal states: froth properties, such as bubble size, bursting gral-Derivative (PID), which has been widely used in many types of
rate, coalescence rate, particle settling velocity, liquid content, processes, including froth flotation. PIDs are often used as regulatory
among others. For intermediate cells: grade, recovery, total controllers and are designed to maintain the most important operating
solids, flowrates, mass pull. variables in their set points. Although PIDs have been considered as a
robust strategy for regulatory control, its performance decreases when
It is worth mentioning that there tend to be discrepancies in the the process is under disturbances that continuously change the optimal
classification of flotation control variables. For example, Lynch et al. operating conditions. This problem arises from the fact that PIDs do not
(1981) considered the reagent addition points and the concentrate explicitly use a process model nor constraints, making it more difficult to
collection points as manipulated variables. These variables, however, adapt to changes. Another disadvantage of PID controllers is their high
are likely part of the circuit design and should not be considered as sensitivity to the interaction between process variables. For processes
manipulated variables for control purposes (Oosthuizen et al., 2017). with very complex dynamics, as is the case of froth flotation, PIDs are
Another discrepancy is that the target for concentrate grades in the in­ not sufficient to maintain the plant in its optimal conditions as it is a
termediate cells, in some cases, could be also considered as a controlled SISO controller (single-input single-output) and, therefore, only one
variable. controlled variable can control one manipulated variable, neglecting the
It must also be noticed that for some researchers, such as Lynch et al. effect of interactions from the other process variables.
(1981) and Hodouin (2011), pulp level set point in each cell or bank is These problems can be addressed by complementing the PIDs (reg­
considered a manipulated variable, while others (Bergh and Yianatos, ulatory control) with advanced control techniques, such as expert con­
1994; Laurila et al., 2002) define the pulp level set point as a controlled trol systems, based on image analysis, Artificial Neural Networks (ANN),
variable, since it changes as a consequence of manipulating the tailings Fuzzy Logic (FL), and Model Predictive Control (MPC), among others. In
flowrate and air flowrate, instead of being directly manipulated. Tailings particular, it has been widely accepted that MPC is one of the advanced
flowrate may be also set as a ratio of the cell feed flowrate and, thus, the control techniques that is capable of dealing with complex processes,
concentrate flowrate can be defined if no wash water is added. A sum­ such as froth flotation, due to its ability to cope with multivariable
mary of the control variables used in the studies discussed in this liter­ systems, considering several process constraints (Desbiens et al., 1994;
ature review manuscript is shown in Table A1, in the Appendix. Desbiens et al., 1998; Bouchard et al., 2009; Sbarbaro and del Villar,
It should also be noted that with improved process instrumentation 2010; Bergh and Yianatos, 2011). The MPC technique is addressed in
some of the control variable categories may change. Problems with the more detail in Section 2.
available instrumentation may be considered as one of the biggest Expert control systems are based on human experience, using heu­
challenges at industrial scale (Laurila et al., 2002; Carr et al., 2009; ristic rules, and knowledge-based or expert opinions. The main idea of
Shean and Cilliers, 2011; Bergh and Yianatos, 2011), and it could this type of controller is to take adequate decisions and control the
directly affect the performance of any type of controller. For example, process by imitating the reasoning from experts (Jovanović and
Bergh and Yianatos (1996, 2003) stated that accurate on-line estimation Miljanović, 2015). Flotation control strategies have been implemented
of concentrate grades using an XRF analyser, generally demands a sig­ as expert systems to control metallurgical objectives (i.e. grade and re­
nificant amount of work in maintenance and calibration. Additionally, covery), via the manipulation of froth depth, air flowrate and wash
Bergh and Yianatos (2003) also highlight that in a number of flotation water flowrate set points (Mckay and Inchausti, 1996; Bergh and Yia­
plants worldwide, the concentrate grade and recovery data usually have natos, 1996; Bergh et al., 1999). Expert controllers are commonly
a large variability. It must be emphasised, however, that this literature implemented as an alternative in the absence of mathematical models

2
P. Quintanilla et al. Minerals Engineering 162 (2021) 106718

Fig. 2. Schematic representation of the hybrid neural modelling technique.

Adapted from Cubillos and Lima (1997). Fig. 3. Schematic representation of fuzzy control (Adapted from Gupta and
Yan (2006)).
required by other types of advanced controllers (Bergh et al., 2013), and
they can also form a higher level control on top of more conventional weights) needed to develop models based on neural networks are few
control strategies. However, these controllers have some disadvantages compared to other systems, there are still issues related to the possibility
as they are based on decision making on local set points between steady of over-fitting due to the many associated degrees of freedom, as well as
states, without considering either the process dynamics or process to the availability of sufficient appropriate training data.
constraints (Bergh and Yianatos, 2003; Bergh et al., 2013). Their effi­ Fuzzy logic (FL) control is based on knowledge gained by experience
ciency is still debatable and their performance could be improved if that is represented by numbers between 0 and 1 (Gupta and Yan, 2006),
robust and sufficient instrumentation were available (Shean and Cilliers, depending on the degree of truth. The relationship between variables, in
2011). terms of process control, can be written using “IF-THEN” logical rules.
Advanced flotation controllers based on froth image analysis have Fig. 3 shows a schematic of the fuzzy control strategy. A number of
been developed in recent years. Machine vision can, for example, different FL based control strategies have been discussed in the literature
determine some characteristics of the froth surface at the top of a (Bergh et al., 1998; Osorio et al., 1999; Hodouin et al., 2001; Liao et al.,
flotation cell, such as the number of bubbles, bubble size, bubble shape, 2011), with these being tested using simulators. A recent example of FL
density, speed and stability (Fu and Aldrich, 2019). Aldrich et al. (2010) control applied in a flotation system is the study carried out by Jahed­
classified process control strategies that are based on image analysis into saravani et al. (2016), in which an FL controller for a batch-flotation
four groups: (i) based on the appearance of the froths to take specific system was designed using empirical data and froth images. Parame­
control actions (Moolman et al., 1994; Moolman et al., 1995a; Moolman ters were calibrated by evaluating the system behaviour at several
et al., 1995b; Moolman et al., 1995c; Moolman et al., 1995d; Moolman discrete operating conditions due to the lack of reliable dynamic flota­
et al., 1996; Cipriano et al., 1998; Holtham and Nguyen, 2002; Wang tion models suitable for control purposes.
et al., 2003; Bartolacci et al., 2006); (ii) based on froth features that While Jahedsaravani et al. (2016) report that satisfactory results
allow operating conditions in the process to be inferred (Niemi et al., were obtained, there still remains a need for an efficient method that can
1997; Hyötyniemi and Ylinen, 2000; Ventura-Medina and Cilliers, 2000; deal with a large number of parameters without the need for these to be
Bonifazi et al., 2001; Citir et al., 2003); (iii) based on the detection of manually tuned (Lewis, 1997), since this makes FL a very time-
rare behaviour in the process and its consequent adjustment by using expensive type of controller (Bergh et al., 1998).
“process maps” or “control charts” that are capable of identifying the To deal with the complex dynamics of the flotation process, a major
nature of the problem (Liu et al., 2005); and (iv) based on the use of focus on model-based predictive controllers (MPC) has been undertaken,
image variables as input for classical control strategies (Bartolacci et al., generating considerable recent research interest. As previously
2006; Liu and MacGregor, 2008; Jahedsaravani et al., 2016; Riquelme mentioned, MPC is ideal for implementing in multivariable processes
et al., 2016). Although image analysis techniques have numerous ad­ such as froth flotation, in which several process constraints must be
vantages such as their consistency across all cells monitored, the pos­ taken into account (Desbiens et al., 1994; Desbiens et al., 1998; Bou­
sibility of linking to plant data, and a high frequency of measurements chard et al., 2009; Sbarbaro and del Villar, 2010; Bergh and Yianatos,
(Forbes, 2007), some features, such as the estimation of mineral con­ 2011). The most crucial part of MPC strategies is the model development
centration in the froth phase have not been fully implemented yet at itself (Desbiens et al., 2000; Maldonado et al., 2009; Sbarbaro and del
industrial scale (Aldrich et al., 2010; Fu and Aldrich, 2019). For a Villar, 2010; Bergh, 2016), which must ensure that models accurately
comprehensive literature review on the topic of online monitoring and describe non-linear processes, such as those occurring in flotation
control for flotation, the reader is referred to Aldrich et al. (2010). (Hodouin et al., 2001). Flotation modelling is a particularly difficult task
An artificial neural network (ANN) is a mathematical structure that is as it is affected by a great number of variables to different extents
used to identify relationships between input and output data that have (Arbiter and Harris, 1962; Laurila et al., 2002), as well as due to its
complex, nonlinear characteristics (Hsu et al., 1995). In the case of complex dynamics caused by the interaction between phenomena in
flotation, models to represent relevant phenomena by using ANN have both the pulp and froth phases (Neethling and Brito-Parada, 2018).
been proposed (Gupta et al., 1999; Al-Thyabat, 2008; Mohanty, 2009; A number of literature reviews focused on flotation control have
Chelgani et al., 2010; Massinaei and Doostmohammadi, 2010). For been published in recent years, though none of these has focused on
flotation control, models based on ANN must be capable of maintaining either model-based controllers or the dynamic modelling required when
operating conditions that yield the desired metallurgical objectives, i.e. implementing such strategies. For example, an overview of process
recovery and/or concentrate grade, through the adequate manipulation control in the mineral processing industry - including key aspects for the
of the system’s set points (Shean and Cilliers, 2011). Gupta and Yan flotation process - was presented by Hodouin (2011); while an overview
(2006), for example, carried out a study based on ANN modelling to of control and simulation implemented specifically in flotation columns
control the collector addition rate in order to maintain the recovery set can be found in Bouchard et al. (2009). In the latter, a sub-classification
point, while Cubillos and Lima (1997), implemented hybrid-neural of flotation models was made according to their use, as those used for (i)
modelling, a schematic of which is shown in Fig. 2, for both adaptive models to predict recovery, (ii) models to study the dynamics of the
identification and optimisation of the flotation system, using models at process, and (iii) models used as soft sensors. Nonetheless, whereas
steady state. Although the number of adjustable parameters (network Bouchard et al. (2009) mentioned the most important models used in

3
P. Quintanilla et al. Minerals Engineering 162 (2021) 106718

flotation column studies, no further analysis of models used for flotation Table 1
control was made. Shean and Cilliers (2011), on the other hand, dis­ Objective functions used in MPC strategies for the flotation process.
cussed advanced control strategies implemented in mechanic flotation Study Objective function
cells. Another literature review on flotation control is that of Jovanović ( )
Zaragoza and J = PCu ⋅RMFCu − Ppen ⋅ gpan − g − PFr ⋅MFr − PCo MCo
et al. (2015), in which the hierarchical levels in flotation control, with an Herbst (1989)
emphasis on model predictive control and intelligent control (expert Perez-Correa J =
∑p ∑C
e k+i
i=1 e k+i Γ̂
′ ′
̂ + i=1 Δuk+i− 1 ΛΔuk+i− 1
control and machine vision), were discussed. et al. (1998)
{ (
Maldonado et al. )2 }
The aim of this manuscript is, therefore, to critically analyse and ∑Ny ∑HP
J = j=1 i=HS wj ⋅ ̂
r j (k + i/k) − ̂
y j (k + i/k) +
(2007, ?)
classify, for the first time in the literature, the existing froth flotation ∑Nu ∑HC − 1 ∑Ny
λp Δup (k + i)2 + j=1 ρj ν2j
models that are amenable to be used for control purposes, as well as to p=1 i=0
( )2
Maldonado et al. ∑N 2
J = Q gCCN − ̂g CC + R j=1 gTCy
highlight the modelling areas that require further research to be effec­
(2007)
tively implemented into MPC strategies. Putz and Cipriano ∑N− 1 ⃦ ⃒ ⃒ ⃦2 2
J = i=0 y4 (k + i + 1 k) − w(k + i k) Q + ‖Δu1 (k + i|k)‖R ) +
⃦ ⃒ ⃒ ⃦
The analysis of the literature reveals that the most commonly used (2015)
‖Δu2 (k + i|k)‖2p + ‖Δu3 (k + i|k)‖2S )
models for MPC are the empirical models, which have demonstrated ⃒ ⃒ ⃒
Tian et al. (2018) J =
∑N− 1 ⃒ ⃒ ⃒
x(z, k + j⃒k), Qx(z, k + j⃒k)〉 + < u(k + j + 1⃒k),
good performance at laboratory and pilot-plant scale. However, the j=0 [<
⃒ ⃒ ⃒
presence of disturbances that continuously change the process condi­ ⃒ ⃒ ⃒
Ru(k + j + 1⃒k) >] + < x(z, k + N⃒k), Qx(z, k + N⃒k) >
tions means that the process can move beyond the conditions over which
the models were originally calibrated. Purely empirical models thus
typically perform badly when extrapolating to predict performance. A control law is obtained by minimising the objective function (Eq. (1)) in
wider range of disturbances can be reliably predicted by implementing order to calculate the values for the control u.
phenomenological models. Most of the existing phenomenological The objective functions of different MPC strategies that have been
models are based on the kinetics of the flotation process, assuming well- developed for flotation are shown in Table 1. As can be seen, these
mixing conditions or plug flow, and have been used to represent both the objective functions can vary from the general form presented in Eq. (1)
pulp and froth phases. While kinetic descriptions provide good ap­ in different ways. For example, Zaragoza and Herbst (1989) used an
proximations for pulp phase behaviour, they are not adequate to model economic approach as objective function. This economic approach
froth phase behaviour. considered the price of the copper (PCu ), a penalty for concentrate grade
The froth phase is neither close to being well mixed nor plug flow less than a constraint value (Ppen ), and the reagent costs (i.e. price for the
conditions, with complex relative motions between phases. This is frother (PFr ) and collector (PCo )). Additionally, their objective function
further complicated by the influence of bubble coalescence and bursting, considered also performance aspects, such as the metallurgical recovery
which is hard to predict due to its complex interrelationship with the (R), the concentrate grade (g), and the operating variables: copper mass
flotation chemistry, particle properties and flow behaviour. The flowrate (MCu ), frother flowrate (MFr ) and collector flowrate (MCo ).
importance of including the froth phase into control strategies is that A structure more similar to that of Eq. (1) is found in Putz and
froth stability is related to the mobility of the froth, and therefore, to the Cipriano (2015). In this case, they considered the final tailings grade
overall solids recovery. Additionally, bubble size in the froth determines (y4 (k)) as output, which must follow a given reference (r(k)). The
the amount of water recovered in the concentrate, which is directly changes in the position of output control valves (Δu1 (k),Δu2 (k),Δu3 (k))
linked to the entrained solids, and thus, to the concentrate grade. was also considered in their objective function. A similar approach was
In the following section, a brief overview of Model Predictive Control taken by Maldonado et al. (2007) and Maldonado et al. (2009). In their
(MPC) is presented for a better understanding of this strategy, as well as objective function, while the tracking error (i.e. the difference between
to provide insight into the importance of modelling for control purposes. the trajectory ( r̂j (k + i/k)) and the predicted controlled variables ( ŷj (k +
i/k) were weighted by the term wj , the control changes on the manip­
2. Model predictive control ulated variables (Δup (k + i)) over the control horizon Hc , were weighted
by the term λp . In addition, Maldonado et al. (2007), Maldonado et al.
Model-based predictive control (MPC) is a wide set of different (2009) added a third term to the objective function (ρj νj ) in order to
strategies that have in common the use of an explicit model of a process “soften the output constraints” (Maciejovski, 2002). In Maldonado et al.
and the minimisation of an objective function (Camacho and Bordons, (2007) the objective function was stated as the minimisation of the
2007) to maintain the controlled variables at the desired set-point with tailings copper grade in each bank. However, although the tailings grade
optimal trajectories. All MPC strategies need to have: (1) a prediction (gTCuj ) is generally easy to measure, it must be noted that maximising
model, (2) an objective function, and (3) a control law. The prediction recovery by reducing the tailings grade could be considered as an
model is, in fact, the most crucial part of MPC. This model must repre­ oversimplification of the economic maximisation calculation.
sent the dynamics of the process as accurately as possible and, at the Camacho and Bordons (2007) describe several advantages that MPC
same time, it must be simple enough to be solved in the shortest possible has over other control strategies:
time.
The general form of the objective function is given in Eq. (1), in • Tuning is reasonably easy, particularly for non-trained personnel,
which the future output (̂ y ) should follow a given reference trajectory since control concepts are very intuitive;
(r), and the necessary control changes (Δu), which can be weighted, as • it can be used in any kind of process, even those with complex dy­
follows: namics or which are unstable;
∑
N2
()
∑
Hc
() • it can be used as a multivariable controller;
J(N1 , N2 , Hc ) = 2
ω j [̂y (t + j|t) − r(t + j)] + λ j [Δu(t + j − 1)]2 , • it has intrinsic delay compensation; and
j=N1 j=1 • plant constraints can be simply included during the design of the
(1) controller.

where N1 and N2 are the minimum and maximum prediction horizons, A reliable dynamic model is needed to implement an MPC strategy.
Hc is the control horizon. The coefficients ω(j) and λ(j) are functions that Dynamic models, of the form presented in Eq. (2), are needed as they
consider the future behaviour. This coefficients can be set to cover a allow the prediction of future plant responses under given operating
wide range of control options, from smoother control to tighter ones conditions (Bergh and Yianatos, 2011):
(Camacho and Bordons, 2007). r(t +j) is the reference trajectory. The

4
P. Quintanilla et al. Minerals Engineering 162 (2021) 106718

• The objective function is optimised by staying as close as possible to

the “reference trajectory” (r(t + j)), which refers to the set point (or
an approximation of it).
• Step 1 is repeated when the next sampling instant y(t +1) is known.
The control signal u(t|t) is sent to the process, while the future con­
trol signals are discarded.

Fig. 4 presents the basic structure of an MPC implementation. While

models of the process are used to predict the future plant outputs based
on the proposed future control signals, the optimisation is done by
minimising a cost function and considering plant constraints based on
manipulated and/or controlled variables (Camacho and Bordons, 2007;
Fig. 4. Basic structure of an MPC implementation. (Adapted from Bor­ Maciejovski, 2002; Rossiter, 2003; Veselý and Rosinová, 2010). The
dons (2000)). control signals are calculated according to the optimisation cost function
and plant constraints; if inequality constraints exist, numerical methods
( )
dx with more calculation load are needed. In general terms, MPC strategies
= f x, u, w, θ, t , (2)
dt differ from each other in the particular models used to describe the
process, how disturbances are estimated, and the cost function used to
where x is the vector of states, f is a set of nonlinear process functions, u optimise the process (Camacho and Bordons, 2007).
is the inputs vector, w is the process noise vector and θ is the model Although MPC has had some successful implementations at indus­
parameters vector (Herbst et al., 1992). In the case of flotation, con­ trial scale, specially in petrochemical plants (Qin and Badgwell, 1997), it
servation of mass usually forms the basis for the dynamic models. has not been effectively implemented in every industry due to the need
Modelling for flotation control is extensively reviewed in Section 3 in to have relatively simple, dynamic models of the process (Bordons,
this paper. 2000) that are accurate enough to represent the process behaviour.
An MPC strategy can be carried out through the implementation of Minerals processing is not an exception in terms of lack of imple­
models to predict the future response of the plant over a finite time mentations of MPC, and it is particularly challenging for froth flotation
horizon. This finite control horizon is used to apply the first control as it has complex dynamics that are difficult to be accurately represented
signal, as well as to repeat each of the following steps, by using new using simple models. While detailed flotation models have been devel­
measured variables (Camacho and Bordons, 2007; Bordons, 2000), as oped for design and analysis purposes, they are not tailored for process
follows: control. The models used for flotation control are classified and analysed
in the following section.
• At each instant t, the outputs of the process are predicted by using the
plant model proposed for a given horizon Hc . 3. Modelling froth flotation for control purposes
• For j = N1 …N2 the predicted outputs ̂ y (t +j|t) (this notation refers to
the value of the variable at instant t + j that is predicted at instant t) Models can be classified in different ways, depending on their
depend on the known values from the past (up to t) and on the future characteristics and purposes. For example, Gharai and Venugopal
control signals u(t +j|t) for j = 0…(Hc − 1) that are calculated and (2016) presented a general classification of the flotation models, iden­
sent to the system. tifying two main groups: micro-scale and macro-scale models. Micro-
scale models have been used to describe the chemical and physical

Fig. 5. Classification of flotation models according to their final purpose. Flotation models for control purposes must be dynamic, and can be sub-classified as
empirical, phenomenological and hybrid (highlighted in solid red lines). (For interpretation of the references to color in this figure legend, the reader is referred to the
web version of this article.)

5
P. Quintanilla et al. Minerals Engineering 162 (2021) 106718

Table 2 postulated that flotation models can be classified as empirical (used for
Summary of models of equations used to represent flotation variables as part of Principal Component Analysis (PCA), multivariate regression, Partial
different type of models discussed in this literature survey. Least Squares (PLS), neural-networks, transfer functions) or phenome­
Variable Equation number nological; steady-state or dynamic; deterministic or stochastic; causal
Empirical Phenomenological model Hybrid
(input-output model) or non-causal (e.g.: mass conservation con­
model model straints); linear or non-linear; based on mathematical equations or fuzzy
rules.
Attached particles in (81)
the froth phase In this paper, models are classified into two main groups according to
Attached particles in (79) their purpose: models developed for (i) simulation purposes (i.e. analysis
the pulp phase and design) or (ii) control purposes, as shown in Fig. 5. While empirical
Attachment rate (27) and phenomenological models can be found for both simulation and
constant in the froth
phase
control purposes, hybrid models have been developed for control pur­
Attachment rate (23) poses only. Despite the fact that a large number of flotation models have
constant in the pulp been developed assuming steady-state, these models have been mostly
phase used for design, simulation and off-line optimisation instead of control
Bias rate (21)
(Bergh and Yianatos, 2011; Bergh et al., 2013).
bubble surface area flux (17)
Bulk fluid velocity due (30) It should be noticed, however, that a strict classification of flotation
to drainage models cannot be proposed, since models can combine aspects of
Collection rate constant (6) different approaches in order to improve their overall utility (Jovanović
Concentrate grade (12) et al., 2015). In fact, although there is not a clear distinction between
Detachment rate (29)
constant in the froth
phenomenological and empirical models, the models presented in this
Detachment rate (28) paper were classified according to the characteristics that represent
constant in the pulp them better. The form of empirical models can be influenced by
Drainage rate constant (4), (7) phenomenological considerations, whereas many of the unknown pa­
Flotation rate constant (3), (5) (34), (35)
rameters within phenomenological models can be obtained by empirical
Gas hold-up (20) (72), (74)
Impeller power (26) means. This paper will only focus on froth flotation models for model-
Liquid holdup (22) based controllers. Reviews on flotation modelling for simulation pur­
Pulp level (39), (42), (65), (66), (67) (70) poses rather than control can be found in Varbanov et al. (1993), Herbst
Sauter mean bubble (19) (18) and Harris (2007), Alves dos Santos et al. (2014), Jovanović et al.
diameter
Solid mass concentrate (10)
(2015), Wang et al. (2015), Gharai and Venugopal (2016), Dinariev and
flowrate Evseev (2018), Prakash et al. (2018), Wang et al. (2018), among others.
Solid mass in the froth (46), (48), (50), (51) Modelling for control purposes is a difficult task since most of the
phase models developed so far have physical parameters that cannot be
Solid mass in the pulp (37), (38), (40), (45), (47),
adequately measured (or even estimated) in a plant. Additionally, the
phase (49), (53), (54)
Solid mass in the (43) complexity of flotation is due to its stochastic behaviour as well as the
tailings lack of reliable instrumentation, which makes it more difficult to
Tailings flowrate (16) (41) develop simple models for control purposes that can be calibrated with
Tailings grade (44) either industrial (Perez-Correa et al., 1998; Casali et al., 2002; Maldo­
Turbulent aggregate (24)
nado et al., 2007; Putz and Cipriano, 2015) or laboratory-scale (Bascur,
velocity
Unattached particles in (80) 1982; Maldonado et al., 2007; Maldonado et al., 2009; Maldonado et al.,
the froth phase 2010; Shean et al., 2017; Shean et al., 2018) data.
Unattached particles in (78) It should also be noted that most of the models that have been
the pulp phase
developed for flotation control have focused on the pulp phase.
Upward gas velocity of (75)
the bubble in the pulp Simplification of the froth phase (or even completely neglecting it) has
Volume of liquid in the (69) been the approach of several authors for flotation process simulations
froth (Bergh et al., 2013). A deeper discussion of the models used for MPC in
Volume of liquid in the (68) froth flotation is presented in the following sections. A summary of the
pulp
models used to represent the each flotation variables mentioned in this
Volumetric concentrate (9) (15)
flowrate literature review manuscript is presented in Table 2.
Water drainage (32)
Water entrainment (31) 3.1. Empirical models
Water in the (33) (58)
concentrate
Empirical models are developed from data analysis through statis­
tical methods that relate input-output measurements from the plant
relationships among sub-processes within a flotation cell (Polat and (Shean and Cilliers, 2011). Multiple linear regression methods or spline
Chander, 2000; Jovanović et al., 2015). Simplifications and combination regression methods are commonly used for such analysis (Mular, 1972;
of micro-scale models form the basis of macro-scale models, which have Whiten, 1972; Lynch et al., 1981), from which parameters with limited
been used for the prediction of the behaviour of entire flotation cells – or physical meaning are obtained (Polat and Chander, 2000).
even banks– the description of process parameters using experimental The collection of data for the development of empirical models can
data, the design of plant layout, and the development of control stra­ be done via online or off-line methods. Online data is obtained via online
tegies (Polat and Chander, 2000; Rojas and Cipriano, 2011; Jovanović instrumentation that allows updating the empirical parameters in the
et al., 2015). Given the massive difference in scale between models for model. Off-line collection methods can be based on either daily oper­
those for an entire cell or banks, it is important to classify models based ating data or from a previously planned campaign (Lynch et al., 1981).
on both their scale as determining their direct utility for MPC. Shorter model development time in comparison with other types of
Another classification was proposed by Hodouin (2011), who models is one of the main advantages of empirical models, specially
when the final purpose of the model that is being developed is previously

6
P. Quintanilla et al. Minerals Engineering 162 (2021) 106718

known. These models are, however, valid only under a certain range of their empirical models, flotation rate constants (Kij ) for each mineral­
operating conditions and must be used for a particular plant, which ogical class i in the flotation bank j, were estimated as a function of froth
treats a particular type of ore, under a well-known operating range. depth (hf ), tailings grade (gT ), tailings volumetric flowrate (QT ) and
Purely empirical models typically decrease in accuracy rapidly if the empirical parameters (p), as shown in Eq. (8). Several sampling tests in a
conditions move outside the calibration range. Although empirical Chilean mine were conducted to validate the proposed models, with the
flotation models have been proposed since the late 60 s by Faulkner model parameters determined by using the standard least squares
(1966),Pitt (1968) and Smith and Lewis (1969), these need to be algorithm.
adjusted for each new mineralogical conditions and plant layout to ( )
which they are to be applied. Kij = f hfj , gTi(j− 1) , QTj− 1 , p . (8)
An example of an empirical model used for flotation control is the
one developed by Perez-Correa et al. (1998). Flotation kinetics (Kp ) and The concentrate volumetric flowrate (QC ) was calculated as a linear
drainage constants (Ke ) were empirically determined, by relating them function of froth height (hf ), as shown in Eq. (9).
to the effect of collector and frother addition rates. As a result, poly­ QCj = c1j − c2j hfj , (9)
nomial relationships (Eqs. (3) and (4)) between these variables were
obtained as follows: where c1j and c2j are empirical parameters.
Even predominantly empirical models can be influenced in their
K ip = ai1 Q3col + ai2 Q2col + ai3 Qcol + κicol,0 , (3)
form by phenomenological considerations, as in Eq. (10), in which the
concentrate mass flowrate (MSCij ) depends on the empirical flotation
K ie = κie0 + bi1 Qcol + bi2 Qf , (4)
constant (Kij , from Eq. (8)). The concentrate mass flowrate was used to
where Qcol is the collector addition rate, and Qf is the frother addition calculate the total solid mass flowrate (MSC , from Eq. (11)) and
rate. The empirical parameters (a1 , a2 , a3 , κcol,0 , κe0 , b1 and b2 ) were concentrate grade (gCij , from Eq. (12)) as follows:
determined for each mineralogical class. In the study by Perez-Correa MSCij = Kij MSPij , (10)
et al. (1998), the mineralogical classes considered were (1) “rich min­
eral”, mainly chalcopyrite, and (2) “poor mineral”, mainly gangue. Since ∑
3
experimental data showed that the amount of iron in the poor classes in MSCj = MSCij , (11)
the feed stream directly affected the rougher concentrate grade, the i=1

flotation rate constant of the poorest mineralogical class was instead

MSC
represented by the Eq. (5). A more comprehensive study would include gCij = 100 ∑3 ij , (12)
more mineralogical classes as the iron can be present in more than one k=1 MSCkj

mineral or as mixed mineral grains, which would further complicate the

where MSPij is the solid mass in the slurry, while the cumulative
model of the flotation rate constant.
( ( )) concentrate solid (MSCij ) mass flowrate was defined as shown in Eq. (13):
K ipf = κip 1 + 0.1 gFe − gFe,0 , (5)
j
∑
MSCCj = MSCk , (13)
where Kipf is the flotation rate of the poorest mineralogical class modified k=1

by iron content, κip is the flotation rate of mineralogical class i from Eq.
and the cumulative concentrate grade (gCCj ) was defined as:
(3), gFe is the iron grade, and gFe,0 is the initial iron grade. It should be
noted that this flotation rate model is rather unconventional, and it was j
∑
not found in any other study analysed in this literature survey. MSC1k
Interestingly, despite the fact that the frother and collector flowrates gCCj = 100 k=1
j ∑
. (14)
∑ 3
were originally defined as manipulated variables by Perez-Correa et al. MSCik
k=1 i=1
(1998), these variables were in the end considered as disturbances in
three of the four simulated cases, and only the pulp level was considered Regarding the estimation of concentrate volumetric flowrate, Putz
as a manipulated variable when implementing the control strategies. and Cipriano (2015) also presented a model that does not only depend
Additionally, the accuracy of the models was not reported as the focus of on the froth height, but also on the pulp level, feed and tailings volu­
the study was to compare different control strategies rather than to metric flowrate, and logic rules 1 (δ1 and δ2 ), as shown in Eqs. (15) and
develop and test control models. (16):
While the models for the above mentioned kinetic constants (i.e. Eqs. ( ) ( )( )
QiC = αiC hiP + hiF − himax δi1 δi2 + QiF − QiT 1 − δi1 , (15)
(3) and (4)) were developed for different mineralogical classes i, Putz
and Cipriano (2015) extended these models to consider not only a
where αiC is a tuning constant, hP is the pulp level, hf is the froth height,
change in each mineralogical class, but also in each cell j, and for each
and hmax is the total flotation cell height. It is important to note that the
granulometry class k, as shown in Eqs. (6) and (7). Neither Perez-Correa
froth height was considered constant due to an absence of advanced
et al. (1998) nor Putz and Cipriano (2015) reported the value of any of
models to estimate it.
the adjustable parameters in the models.
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
K ijk ijk 3 ijk 2 ijk ijk
col = a1 Qcol + a2 Qcol + a3 Qcol + κcol,0 , (6) QiT = αiT ui hiP − hi+1 P + Δh ,
i (16)

K ijk ijk ijk ijk

e = b1 Qcol + b2 Qf + κe0 , (7) where αiT is a tuning constant, ui is the control valve position, and Δh is
the physical height difference of two consecutive flotation cells. It
ijk
where Kcol is the collection rate constant, and Kijk should be noted that there is usually a minor difference in height be­
e is the drainage con­
stant. Flotation rate constants were also modelled by Maldonado et al. tween cells, but a significant difference in pulp level between cells as the
(2007), for a rougher flotation control strategy based on dynamic pro­ mass is removed down a flotation bank. This model is based on Torri
gramming (Bertsekas, 1995). In their implementation, both phenome­
nological and empirical models were used. The phenomenological
models from their work can be found in Section 3.2, Eqs. (40)–(44). For 1
Logic rules are part of hybrid models, which are addressed in Section 3.3.

7
P. Quintanilla et al. Minerals Engineering 162 (2021) 106718

Table 3 shown in Eq. (21), as a function of the fraction of wash water, which is
Description of operational constraints proposed by Maldonado et al. (2009). also calculated by using conductivity (Eq. (22)). In flotation columns,
Constraint Description wash-water rate (Jw ) is a crucial operating variable as it is directly linked
to froth stability and it is used to enhance drainage, thus reducing
Jg ⩽Jgmax To prevent: hydraulic entrainment, loss of the interface and froth
“burping” entrainment. This is an important variable to control in flotation col­
Jg ⩾Jgmin To keep solid in suspension umns since a very high wash-water rate can lead to inefficient froth
Jw ⩽Jwmax To avoid: froth mixing and wash-water short circuiting cleaning action due to increasing in froth mixing and water short-
Jw ⩾Jwmin To promote froth stability and to facilitate transfer of collected particles circuiting to the concentrate, limiting the overall metallurgical perfor­
into the concentrate mance (Yianatos and Bergh, 1995). Table 3 shows the six inequality
JB ⩽JBmax To avoid reduction of collection residence time for valuable minerals equations that were implemented by Maldonado et al. (2009).
JB ⩾JBmin To perform cleaning action thus reducing gangue entrainment
JB = 0.003966⋅εw − 0.03409, (21)

celli’s principle and depends on the position of the output control valve The liquid hold-up (εw ) is calculated as:
(ui ∈ {0, 1}). The Torricelli’s principle relates the velocity of fluid ( )
k − k*
leaving a cylinder (as the flotation tank) to the height of the fluid. In this εw = 100 f , (22)
kf − kw
context, Torricelli’s principle was used to develop a model for the tail­
ings flowrate in terms of the pulp height (or “fluid height”) and the where kf is the conductivity of the water in the feed, k* is the conduc­
control valve. tivity of the interface, and kw is the conductivity of wash-water.
A nonlinear controller for the bubble surface area flux was designed Although the use of sensors based on conductivity was not of major
by Maldonado et al. (2010). The controlled variable was the bubble concern to Maldonado et al. (2009) since these sensors have been
Sauter diameter and the manipulated variable was the superficial water satisfactorily tested on plants (Gomez et al., 2003; Bartolacci et al.,
velocity that passes through a sparger ring that was installed in the 2008), differences in the estimation of bias rate were found when
experimental rig, which allowed to control the bubble size indepen­ comparing these measurements to a water balance in the collection
dently from gas velocity. The sparger ring used was a “frit-and-sleeve” zone. This suggests that either the conductivity measurements have an
sparger, specially designed by Kracht et al. (2008), which uses a porous intrinsic error that should be considered when controlling a flotation
ring and a sleeve to control the bubble size, as described in Maldonado column using the proposed models, or that the model relating liquid
et al. (2010). The bubble surface area flux (Sb ) can be calculated as Eq. hold-up to flowrate is the source of the error.
(17), and it can be controlled by manipulating the ratio between su­ The fact that froth height has been assumed constant in most models
perficial gas velocity (Jg ) and the Sauter mean bubble diameter (d32 ): for flotation control indicates there is an opportunity to enhance these
6Jg models so that they also take into account all the phenomena that occur
Sb = , (17) in the froth phase. For example, particle attachment and detachment
d32
processes in the froth phase have been considered in some studies, such
where d32 is calculated as: as Zaragoza and Herbst (1989), in which a hierarchical control based on
∑Nb 3 a simplified version of the models proposed by Bascur (1982) was
dBPi implemented. In order to simplify those models, Zaragoza and Herbst
d32 = ∑i=1
Nb 2 , (18)
i=1 d BPi (1989) assumed that the attachment and detachment processes are at
equilibrium in the pulp and the froth phases. This assumption is ques­
where dBP is the bubble diameter in the pulp phase, and Nb is the number tionable in terms of application for control since this equilibrium may be
of bubble size classes. affected by dynamic changes in, for example, particle size, head grade in
As can be seen, Eqs. (17) and (18) are not explicitly time dependant, the feed, and particle liberation.
and therefore, an online estimation of the Sauter mean bubble diameter, The rate constants of the attachment phenomena in the pulp phase
previously proposed by Maldonado et al. (2008), was implemented. A (KPAT
ij ), can be modelled as shown in Eq. (23) (Bascur, 1982):
nonlinear steady-state relationship between Sauter mean diameter (d32 ),
and superficial water velocity (Jls ) was obtained, as shown in Eq. (19). vBP d2iB ut
K PAT
ij = κPAT
j , (23)
d3BP
d32 = 3.706⋅J −ls 0.256 − 0.226. (19)

The strong influence of frother concentration on bubble size was also where κPAT
j is a constant which is determined experimentally, vBP is the
subject of flotation control research by Maldonado (2010), who volume of bubbles in the pulp phase, diB is the mean size of the aggregate
considered frother concentration as an unmeasured disturbance that particle-bubble of size i, and dBP is the bubble diameter in the pulp
directly affects flotation performance due to its impact on bubble size phase. The turbulent aggregate velocity (ut ) can be estimated using Eq.
and froth stability. Maldonado et al. (2009) used electrical conductivity (24) (Schubert, 2008):
to calculate the gas hold-up (εg ) in a laboratory scale flotation column, as ( )2/3
√̅̅̅̅̅̅̅̅̅ ε4/9 d7/9
shown in Eq. (20). Froth height was also considered constant for the Δvt′ 2 ≈ 0.33 1/3
p Δρ
. (24)
dynamic models used to implement the control strategy. vF ρF
( )
εg = 100
kl − kl− g
, (20) where ∊ is the average energy dissipation, dp is the particle size, vF is the
kl + 0.5kl− g kinematic fluid viscosity, Δρ is the density difference between particle
and fluid, and ρF is the fluid density.
where kl and kl− g are the conductivity for the liquid and liquid-gas
The average energy dissipation (∊) in a flotation tank is given by Eq.
mixture, respectively.
(25):
The constrained controller implemented in Maldonado et al. (2009)
was used to minimise the tracking error of the gas hold-up (εg ) and bias ∊=
Pg
, (25)
rate (JB ), which is defined as the net downward water flowrate (i.e. bias VLP
rate = tailings flowrate (water) - feed flowrate (water)) (Del Villar et al.,
1999). The bias rate (JB ) was calculated through the empirical equation where Pg is the power drawn by the impeller, and VLP is the volume of
liquid in the pulp. The power drawn by the impeller can be obtained

8
P. Quintanilla et al. Minerals Engineering 162 (2021) 106718

from dimensional analysis, as in Eq. (26):

( )e2 ( 2 3 )e3
Pg QA N I D I ρL
= e1 , (26)
P0 NI DI σ

where P0 is the power input to impeller when no air is added, DI is the

impeller diameter and NI is its speed, QA is the air flowrate, ρl is the
density of water and σ is the surface tension. e1 , e2 and e3 are constants.
The rate constant of the attachment process in the froth phase (KFAT ij )
is modelled as shown in Eq. (27). This model is a function of collision Fig. 6. Transfer of material between pulp and froth regions in a flotation
frequency (ddBFP ), and the number of bubbles across the height of the froth process (Adapted from Lynch et al. (1981)).
(Bascur, 1982).
( )( ) size, reagents, among many other variables, are common and have a
dp hf
K FAT
ij = κ FAT
j Q A , (27) large impact in the operation at industrial scale, and should therefore be
dBF dBF
considered when modelling for control.
One solution to these model mismatches is the development of
where κFAT
j a constant which is determined experimentally, QA is the
phenomenological models for control. Phenomenological models are
volumetric air flowrate in the pulp, dp is the particle size, dBF is the
based on fundamental aspects of the process (e.g. physics or surface
bubble size in the froth, and hf is the froth height.
chemistry), and are able to predict changes in operating conditions,
The constants rate of detachment phenomena in both pulp and froth achieving the robustness necessary to be applied in MPC strategies
phases were also determined by empirical equations. In the pulp phase, without the need for significant additional model development.
the rate of detachment (KPDT
ij ) is given by Eq. (28). Phenomenological models for flotation control are the focus on the next
section.
K PDT
ij = κPDT
j d2P ut , (28)

where dp is the particle size and ut is the turbulent velocity that can be 3.2. Phenomenological models
calculated as Eq. (24). In the froth phase, the rate constant of detach­
ment (KFDT
ij ) is given by Eq. (29): Unlike empirical models, a phenomenological model is derived from
fundamental laws. While phenomenological models for simulation or
K FDT = κFDT ρ n
j d P u∞ , (29)
ij j analysis purposes have often been sub-classified as kinetic, population
balance and probabilistic (Polat and Chander, 2000; Gharai and Ven­
where ρj represents the specific gravity of species j, n is an adjustable
ugopal, 2016), the latter are not found in the literature pertaining to
constant, and u∞ is the bulk fluid velocity due to drainage, which can be flotation predictive control. In this paper, instead, models are classified
approximated by Eq. (30): as kinetic, population balance, and hydraulic, all of which are discussed
Q below.
u∞ = ( R ), (30) It must be taken into account that while several phenomenological
A3 1 − εg
flotation models already exist, most of them have been used for design or
where QR represents the volumetric drainage rate and A3 (1 − εg ) repre­ analysis purposes since they are often not simple enough to be imple­
sents the effective cross sectional area of the liquid phase in the froth. mented into control strategies. Additionally, models for design and
The population balance approach needs to be complemented with analysis purposes can be steady-state, whereas MPC requires dynamic
hydraulic models as the liquid distribution between the pulp and froth is models.
required. For this, Bascur (1982) proposed empirical models that allow
the estimation of water entrainment (QE ), water drainage (QR ) and 3.2.1. Kinetic models
water reporting to the concentrate (QC ), which can be calculated by Eqs. Kinetic models are developed by considering processes as analogous
(31)–(33), respectively. to a chemical reaction. Lynch et al. (1981) described this “reaction” in
flotation by considering two types of collisions: (i) between molecules
QA and (ii) between hydrophobic particles and air bubbles in the pulp.
QE = a4 , (31)
d0.75
BP Simple kinetic descriptions, however, may ignore the contribution of the
( )0.53 froth to the overall transport of both the valuable material and the
QA
V 0.56 0.4 gangue. These kinetic descriptions can therefore be extended to include
LF A
the transfer of material not just out of the pulp phase, but also into and
A
QR = a5 , (32)
σ0.24 d1.92
BF out of the froth phase. A schematic of particles transfer from pulp to
froth or vice versa is shown in Fig. 6. This schematic represents particle
( )1.5 ( )
QC = a6 L hf − hmax 1 − εg , (33) transfer from the pulp phase due to selective attachment or non-selective
entrainment, as well as the particle transfer from the froth phase to the
where a4 , a5 and a6 are adjustable parameters, L is the overflowing lip pulp phase due to drainage.
length of the cell, hf is the froth height, hmax is the total flotation cell By considering flotation as a chemical reaction, a kinetics model can
height, σ is the surface tension, A is the cross-sectional area of the froth be derived, as shown in Eq. (34) (Garcia-Zuniga, 1935; Arbiter and
volume, and εg is the gas hold-up. Plant data were used in order to es­ Harris, 1962; Gharai and Venugopal, 2016).
timate the parameters required in the aforementioned equations (Bas­
dN
cur, 1982; Zaragoza and Herbst, 1989). = − Kn N n N mb , (34)
dt
One major drawback of empirical models is that they are no useful
for operating conditions other than those used to develop the model. The where K is the flotation rate constant, N is particle concentration, Nb is
operating conditions in an industrial process, however, usually change, the bubble concentration; n and m are their respective kinetic orders.
resulting in empirical models having limited accuracy and robustness. In The negative sign is due to the fact that concentration decreases as
particular for the flotation process, changes in feed mineralogy, particle particles leave the flotation cell.

9
P. Quintanilla et al. Minerals Engineering 162 (2021) 106718

Generally, flotation kinetics are considered to be first-order (Gharai in the pulp was calculated using the Eq. (40), while Eq. (41) was used to
and Venugopal, 2016), as shown in Eq. (35) (Kelsall, 1961; Arbiter and describe the tailings volumetric flowrate. It must be noticed that the
Harris, 1962; Yoon and Mao, 1996), with the bubble concentration de­ tailings volumetric flowrate can also be calculated empirically, as pre­
pendency absorbed into the rate constant: viously shown in Eq. (7). The models included the time variation for
each mineralogical class i in cell j of the solid mass Mij in the pulp phase,
dN
= − KN. (35) and the pulp level hpj in each cell.
dt
Nguyen et al. (1998) developed a first order kinetics model consid­ MSTj− 1 gTi(j− 1)
MSPij = ( ), (40)
ering three-components according to their floatability: non-floating, QT j
̃
kij +
slow and fast, as shown in Eq. (36): Ahpj

( )
Rec = 1 − mf exp − Kf t − ms exp( − Ks t) − mn , (36)
where MSTj− 1 is the solid mass flowrate in the tailings, gTi(j− 1) is the
where mn , ms and mf are the mass fractions of non-floating, slow and fast metallurgical grade in the tailings.
floating species, respectively; and Kf and Ks are the respective flotation QTj = QTj− 1 − QCj , (41)
rate constants. More recently, Neethling et al. (2019) have shown that
flotation cells can sometimes experience zero-order kinetics, and Froth depth, tailings solid mass flowrate and tailings grade were also
explored the cases when this can happen, carrying out modelling and implemented into the MPC strategy tested by Maldonado et al. (2007),
validation. In order to model the transition between first and zero order by using phenomenological models as follows:
kinetics, assuming that the cell is well mixed, the “reaction” is consid­
ered to be between the particles and the available bubble surface area • The froth depth (hfj ) for each cell was calculated as the difference
flux, the flotation rate was defined as Eq. (37): between the cell total height (H) and pulp level (hpj ), as shown in Eq.
(
εg
) (42):
Mi = Ki Cp,i εg χ b 1 − Vcell , (37) ( ) ( )
εg,max
hfj = H − hpj ⋅1000 mm . (42)
where Mi is the solid mass of the specie i, Ki is the flotation rate constant,
Cp,i is the solid concentration, εg is the gas hold-up, χ b is the bubble
surface area, and Vcell is the cell volume. • The tailings solid mass flowrate for each mineralogical class i in the
Although this approach that accounts for this transition kinetics has bank j (MSTij ) was calculated using the Eq. (43):
not been implemented into any control strategy yet, it is amenable to be QTj
tested in future studies. Even though numerous kinetic models have MSTij = MSPij , (43)
Ahpj
been developed for froth flotation analysis and design, as discussed in
Gharai and Venugopal (2016) and Mesa and Brito-Parada (2019), it is where MSPij is the solid mass in the pulp phase.
worth mentioning only those kinetic models used for control purposes • The tailings grade for each mineralogical class i in the flotation bank j
are discussed within this section. Therefore, some of the key aspects of (gTij ) was calculated by assuming perfect mixing in the flotation cells,
kinetic models – such as K distribution, for example – are not discussed
using the Eq. (44):
in this paper.
In terms of flotation control, a study on the optimal control of a MSPij
gTij = 100 . (44)
rougher flotation circuit can be found in Maldonado et al. (2007), based ∑
3
MSPkj
on a dynamic simulator developed by Casali et al. (2002). The models k=1
developed by Casali et al. (2002) are shown in Eqs.(38) and (39). Mass
balance models for the pulp phase of a nine-cells rougher flotation cir­
cuit of sulphide copper ore were implemented, assuming five mineral
classes, perfect mixing, and constant air flowrate. In the same study (Maldonado et al. (2007)) empirical models were
[ ] also implemented, as was previously shown in Eqs. (8)–(14) (Section
dMij
= − Kij + (
QTj
) Mij + MSij , 3.1). While Casali et al. (2002) and Maldonado et al. (2007) have
dt 1 − εg AN j considered the effect of gas hold-up (εg ) in their mass balances (Eqs.
(38) (38)), Putz and Cipriano (2015) developed mass balances that consid­
i = 1, …, 5j = 1, …, 9,
ered the attachment and detachment processes rather than gas hold-up,
as shown in Eqs. (45) and (46). In this case, i represents the cell number
of the bank, j represents the mineral class (defined as high or low) and k
where Mij is the solid mass of the specie i of the cell j, Kij is the flotation
represents the different granulometries in the ore.
rate constants, QTj is the tailings volumetric flowrate in the cell j, εg is the
[ ]
gas hold-up, A is the cross-sectional area of the cell, Nj is the particle dM ijk QiT
(45)
p
= M ijk ijk ijk
f + K e Mf − K ijk
P + i MP ,
ijk
concentration in the cell j, and MSij is the solid mass flowrate. The pulp dt VP
level hP of the cell j is calculated by Eq. (39):
[ ]
dM ijk QiC ijk
(46)
dhPj QFj − QCj − QTj f
= K ijk ijk
P MP − K ijk
e + Mf ,
= ( ) , j = 1, …, 9, (39) dt V iF
dt 1 − εg A
ijk
where QFj is the feed volumetric flowrate to the cell j, QCj is the where Mijk
p is the mass of solids in the pulp phase and Mf is the mass of
concentrate volumetric flowrate from the cell j, and QTj is the tailings solids in the froth phase. Kijk ijk
e and Kp are the respective flotation rates
volumetric flowrate from the cell j. constants, QT is the tailings flowrate from the cell i, V iP and V iF are the
Similarly to Casali et al. (2002), Maldonado et al. (2007) imple­ pulp and froth volume of the cell i, respectively. The collection and
mented mass balance models for the pulp phase, but only three miner­ drainage rate were determined empirically, as shown previously in Eqs.
alogical classes and five flotation cells were considered. The solid mass (6) and (7).

10
P. Quintanilla et al. Minerals Engineering 162 (2021) 106718

Another kinetic balance approach that considers the attachment and from the upward water flow to the babble; and the term βCFa represents
detachment processes is that in Zaragoza and Herbst (1989). In this the particles detachment from the bubble.
work, simplified models for the pulp (Eq. (47)) and froth phase (Eq. The initial conditions for the collection zone models used by Tian
(48)) were proposed in order to implement an advanced model-based et al. (2018) were:
control in a 2-cells flotation circuit.
Ca (0) = Ca0 ,
dM p
(
1− εg p
)
Mp KR QR Mf (52)
Cw (0) = Cw0 ,
= MFD − QT +QE +QA α ⋅( ) + ⋅( ), (47)
dt εg 1+ αp VLP VLF 1+ αf
while for the pulp phase, two cases were considered: downward water
dM f ( ( )) Mf motion and upward water motion, as shown in Eqs. (53) and (54),
= − QR KR + Qc 1 + αf ( )
dt 1 + αf VLF respectively:
( ) (48) ( ) [( ) ]
1 − εg Mp ∂ εwd Cwd (z, t) ∂ Uwd + εwd Us Cwd (z, t)
+ QT + QA αp ( ) , = − α1 AvfCwd (z, t)
εg 1 + αp VLP ∂t ∂z (53)
where Mp and Mf are the solid mass in the pulp and froth, respectively. +ρ1 Cwu (z, t) + kβCFa (z, t),
MFD is the mass flowrate in the feed, QT is the tailings flowrate, QE is the
entrainment water flowrate, QA is the air flowrate, εg is the gas hold-up, ∂(εwu Cwu (z, t) )
=−
∂(Uwu Cwu (z, t) )
− σ1 AvfCwu (z, t) − ρ1 Cwu (z, t)
αp and αf are the equilibrium constant between the attachment and ∂t ∂z (54)
detachment in the pulp and froth phases, respectively. VLP is the volume
+(1 − k)βCFa (z, t),
of the liquid in the pulp, VLF is the volume of the liquid in the froth, and
QR is the water flowrate draining back. where the term ρ1 Cwu represents the transfer of particles from the up­
These models were also coupled with an overall volume balance in ward water flow to the downward water flow.
the flotation cells, as shown in Eq. (68) and Eq. (69), in the next section. The boundary and initial condition for the PDEs were:
Another similar approach, i.e. considering attachment and detachment
processes in the pulp and froth phases, was proposed by Tian et al. CFa (0, t) = Ca (t),
(55)
(2018), where dynamic models were implemented into the model pre­ Cwu (0,t)=Cw (t),
Cwd (h,t)=0,
dictive control strategy for a 2-phase flotation column. These models
have one spatial dimension in addition to the temporal dimension for CFa (z, 0) = fa (z),
which they were also solved. The pulp and froth phases were modelled Cwu (z,0)=fwu (z), (56)
via mass balances, considering the mass concentration of solid particles Cwd (z,0)=fwd (z).

(free mineral, locked and gangue) within the “air phase” and “water A discrete controller was designed by Tian et al. (2018) in order to
phase”. Ordinary differential equations (ODEs) were used to model the implement it as part of a flotation column control system. The feed
solid particles in the pulp phase, as shown in Eqs. (49) and (50): flowrate was used as the manipulated variable while the concentrate
( ( ))
d εg VCa t
() () () grade was used as the controlled variable. Nonetheless, it has to be
= α1 Avf εw VCw t − βεg VCa t − QA Ca t , (49) emphasised that this is not a realistic case scenario as in most flotation
dt
cells the feed flowrate is a known disturbance rather than a manipulated
d(εw VCw (t)) ( ) ( ) variable. To carry out this strategy, the Cayley-Tustin time discretisation
= − α1 Avf εw VCw t + βεg VCa t transformation (Humaloja and Dubljevic, 2018) was used to discretise
dt (50)
( ) ( ) ( ) the coupled ODE-PDEs. A large number of parameters and variables
+QF CF − QT Cw t + Qwd Cwd 0, t − Qwu Cw t ,
were considered as constants by Tian et al. (2018) to evaluate their
proposed MPC strategy.
where εg is the gas hold-up, V is the volume of the pulp phase in a
Among the variables considered as constants were the height of froth
flotation column, Ca , Cw , Cwd , Cwu are the mass concentration of solid
phase and collection region, the air hold-up, air flowrate, attachment
particles in the air phase, water phase, downward water phase, and
and detachment rate parameters (for a full list and the values used for
upward water phase, respectively. α1 is the attachment rate constant and
constants, the reader is referred to Table 1 in Tian et al. (2018), although
β is the detachment rate parameter, Av is the air-water interfacial area
no details were provided for the values given to each parameter or
per unit volume of the flotation column, f is the fractional free surface
variables). Assuming key variables remain constant is far from optimal,
area of the bubbles, εw is the hold-up of the water phase, QA is the air
as changes can have a significant impact on operating conditions. For
flowrate, QF is the feed flowrate, QT is the tailings flowrate, Qwu is the
example, while gas hold-up was assumed to be constant in both the pulp
upward water flowrate, and Qwd is the downward water flowrate.
and froth phases by Tian et al. (2018), this is a variable that is not only
The froth phase was modelled by using Partial Differential Equations
prone to change but also one that is related to pulp height (Shean et al.,
(PDEs), considering the froth as a Plug Flow Reactor (PFR). The PFR
2018), as discussed in Section 3.2.2, and thus has a direct impact on the
assumption considers that the froth phase is not mixed in the flow di­
overall performance of the flotation system. Further work is therefore
rection, but is perfectly mixed in the direction perpendicular to the flow.
needed to improve the models, so that the impact of changes in key
However, it should be noted that modelling the pulp as a plug flow
parameters, such as froth height and gas hold-up, can be predicted.
reactor is only applicable to a column cell and, even there, it ignores the
While Zaragoza and Herbst (1989) also considered attachment and
substantial vertical mixing which may occur in these systems. The solid
detachment processes in their balances, Tian et al. (2018) developed a
mass balance for the air phase was modelled as shown in Eq. (51):
model that was not only a function of time but also of space (1-dimen­
( ) ( )
∂ εg CFa (z, t) ∂ Ua CFa (z, t) sional). Since the latter considered ODEs to represent the pulp phase
=− + α1 AvfCwd (z, t) + σ 1 AvfCwu (z, t) − βCFa (z, t), (only time dependent) and PDEs to represent the froth phase (time and
∂t ∂z
(51) space dependent), Eqs. (49)–(54) were coupled, with the ODE system
providing boundary conditions for the PDE system. Although both
where Ua is the velocity of particles within the air phase. The term Zaragoza and Herbst (1989) and Tian et al. (2018) attempted to inte­
αAv fCwd represents the transfer of particles from the downward water grate the phenomenology of the froth phase into control strategies, their
flow to the bubble; the term σ1 Av fCwu represents the transfer of particles approach was based on kinetic models, rather than the physics of the

11
P. Quintanilla et al. Minerals Engineering 162 (2021) 106718

process. While kinetics can represent the pulp phase, the froth phase is flotation control is presented below.
significantly more complex, and it is dominated by phenomena such as Water recovery: The simplified models developed in Neethling et al.
bubble coalescence and bubble bursting. Kinetic models are not there­ (2003) are focused on predicting the amount of liquid collected in the
fore sufficient to fully represent this phase. Enhancement of flotation concentrate (i.e. water recovery), which is directly related to the amount
control can be achieved by including more of the important phenomena of gangue collected and, therefore, to the grade of valuable ore in the
involved in the process and, for this reason, including froth physics as­ concentrate. To do so, a model that relates the amount of water collected
pects is essential to develop advanced control strategies. with the air flowrate, air recovery and bubble size was developed:
The dynamic of the flotation froths has a strong influence on the ( )
overall performance of flotation cells, and thus, its control is difficult to 1 AJ 2g λ
if α < : Ql = 1− α α
achieve (Neethling and Brito-Parada, 2018). In addition to those con­ 2 k1
(58)
trollers based on froth images (such as those mentioned in the intro­
AJ 2g λ
duction), only a few researchers have addressed the integration of froth 1
if α⩾ : Ql = ,
performance into control strategies. One example of this is the study 2 4k1
carried out by Shean et al. (2017), in which an optimisation system was
where α is the air recovery (in most rougher and scavenger cells α is less
implemented at laboratory scale, based on the Generating Set Search
than 0.5), Ql is the upwards liquid flowrate in the froth, A is the flotation
(GSS) algorithm. This control algorithm was based on maximising the
cell cross-sectional area, Jg is the superficial gas velocity, and λ is the
air recovery – a measurement of froth stability – which was developed
length of the Plateau borders per volume of froth, which is related with
and validated in a flotation tank at laboratory scale. Air recovery is
bubble size as shown in Eq. (59):
defined as the fraction of air entering a flotation cell that overflow as
unburst bubbles (Neethling and Cilliers, 2008). Air recovery can be 1
λ∝ , (59)
calculated at steady-state as: d2BF
vf hf Lη
α= , (57) where dBF is the bubble size in the froth phase. The constant k1 is a result
QA
of the force balance between gravity and viscosity, and it is defined as
where vf is the overflowing froth velocity, which is usually measured Eq. (60):
through image analysis; hf is the height of the overflowing froth over the ρp g
k1 = , (60)
lip, L is the lip length, η is the fraction of air in the froth, which is usually 3CPB μ
close to 1, and QA is the air flowrate into the flotation cell.
It has been demonstrated that a peak in air recovery (PAR) is linked where ρ is the pulp density, g is gravity, Cpb is the drag coefficient, which
to improvements in metallurgical performance (Hadler and Cilliers, is a function of interfacial mobility, taking a value of 49 for immobile
2009; Hadler et al., 2010). For this reason, Shean et al. (2017) consid­ interfaces, which will be an appropriate value for many particle laden
ered PAR as the objective in the optimisation problem in order to find flotation froths, and μ is the pulp viscosity.
the best operating condition (in this study, air flowrate) for the flotation Entrainment factor: The performance of the froth flotation process is
tank. However, it should be noted that this PAR optimisation was only partly determined by the concentrate grade achieved. This is directly
tested for a single flotation tank, which is a case far from optimal as related to the entrainment of gangue material in the concentrate. It has
flotation tanks at industrial scale are generally connected in series, and been demonstrated that the amount of gangue entrained is proportional
therefore, the performance of tanks down the bank is directly affected by to the water recovery presented in Eq. (58). This proportionality is
the performance of those upstream. It should be also noted that the known as the entrainment factor. Neethling and Cilliers (2009) devel­
optimisation strategy was tested in a closed loop, in which the only oped a simplified model for the entrainment factor as a function of
manipulated variable was the air flowrate, maintaining reagent condi­ operating conditions, such as froth depth and air rate:
tions constant and pulp flowrates equal to zero, which is clearly not ( )
comparable with flotation at industrial scale. 1 v1.5
set hf
if α < : Ent ≈ exp − √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
( )
As flotation froths play an important role in the flotation perfor­ 2 D J α 1− α
Axial g
mance, much research has focused on modelling the phenomena , (61)
( )
occurring in this phase, specially focused on simulation purposes (i.e. 1 2v1.5
set hf
if α⩾ : Ent ≈ exp − √̅̅̅̅̅
analysis and design).For example, foam physics models developed by 2 DAxial Jg
Verbist et al. (1996), Leonard and Lemlich (1965) have been further
extended to describe flotation froths (Neethling et al., 2000; Neethling where α is the air recovery, vset is the settling velocity, hf is the froth
and Cilliers, 2002; Neethling et al., 2002; Neethling et al., 2003; depth, Daxial is the axial dispersion, and Jg is the superficial gas velocity,
Neethling and Cilliers, 2009). These models aim to describe both in­ i.e. Qair /Acell . The settling velocity, vset , is assumed constant since the
ternal behaviour and overall performance. solid concentration is generally low. This velocity can be calculated as in
It must be noted that some of these models are not directly applicable Eq. (62). The axial dispersion, Daxial , is calculated as in Eq. (63).
to flotation control as they take the form of PDEs and are too complex to
solve for control purposes. However, they have also been simplified to
2
1 g(ρS − ρl )dp
vset ≈ , (62)
an extent that could make them amenable to control, even though they 3 18μ
were developed assuming steady state. For example, in Oosthuizen and
Craig (2019), the potential use of non-linear flotation models for control J 1.5
(63)
g
DAxial ≈ √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
/ )̅ ,
has been investigated and demonstrated. The benefit of such an ( √̅̅̅
k1 3 − π 2 Pe
approach lies in that the concept of PAR implies that there is an optimal
operating point as recovery goes through a maximum when air rate is
where Pe is the Péclet number, which can be assumed 0.15 (Neethling
varied (i.e. a non-linear model is required). A brief introduction to some
and Cilliers, 2009).
simplified froth models that could be used in future MPC studies for

12
P. Quintanilla et al. Minerals Engineering 162 (2021) 106718

Fig. 7. Schematic diagram of flotation cells. Adapted from Jämsä-Jounela et al. (2003).

The entrainment factor is important to predict the entrained solids single-output) controller and three different MIMO (multi-input multi-
recovery from the estimated water recovery (from Eq. (58)). In Neeth­ output) controllers were tested. In order to determine which type of
ling and Cilliers (2009), it was demonstrated that while particle size has controller is more suitable for pulp levels, simulations in Matlab
a strong influence on the entrainment factor, there is no direct depen­ (Simulink) were performed, finding that SISO controllers are consider­
dence of it on the overflowing bubble sizes. ably less efficient in their ability to control the cell levels than MIMO
Froth recovery: Froth recovery is defined as the fraction of the ma­ controllers. Dynamic phenomenological models to predict the pulp
terial that enters the froth attached to the bubbles that reports to the levels in the cells, as shown in Eq. (65) for the first cell, Eq. (66) for the
concentrate, rather than dropping back into the pulp (Finch and Dobby, 2nd to (n − 1) cells and Eq. (67) for the last one, were proposed. These
1991; Neethling and Cilliers, 2008). By including the froth recovery into equations are valid when there is a tailings valve for each tank, as shown
control strategies, the performance of predictive control strategies could in Fig. 7, and when the cross-sectional area of the cell is constant.
be enhanced as it is related to the particle detachment, as well as the √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
behaviour of unattached particles. Froth recovery could be included into dhp1 q − Kv Cv (u1 ) hp1 − hp2 + Δhp1
= , (65)
control strategies by using a simplified model developed by Neethling dt A1
and Cilliers (2008). This simple froth recovery model is a function of ( )√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ ( )√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
operating variables, such as superficial gas velocity, Jg , air recovery, α, dhpj KVj− 1 Cv uj− 1 hpj− 1 − hj + Δhpj− 1 KV Cv uj hpj − hpj+1 + Δhpj
= − ,
and bubble sizes, as follows: dt Aj Aj
( ( ) )f ( ) (66)
1 α 1 − α Jg 2 db,int f
if α < : RFroth ≈
2 vset dBF,out ( )√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
, (64) dhpn KVj− 1 Cv uj− 1 hpj− 1 − hn + Δhpj− 1 KV Cv (un ) hpn + Δhpn
( )2f ( )f = − ,
1 Jg db,int dt An An
if α⩾ : RFroth ≈
2 4vset dBF,out (67)

where f is the fraction of material that becomes detached from the where hp1 ,hpj , and hpn are the pulp level in the flotation cell number 1, 2
bubble interface during a coalescence event, db,int is the bubble size in to (n − 1) and n, respectively. KV is a proportional constant, Cv is the
the pulp-froth interface, and dBF,out is the bubble size in the froth phase, valve coefficient, and u1 , uj and un are the control valve positions for the
overflowing the cell lip. tailings flowrate from the cells 1, 2 to (n − 1) and n, respectively. Δhp is
Although there have been attempts to develop simple froth models, the physical difference between two consecutive cells.
such as those presented above, there is still a gap in our knowledge on Basic PI-controllers for level in every cell were used to evaluate the
how to include this type of simplified models into predictive control control strategies proposed, and only the feed flowrate was measured
strategies. It should be possible to include these simplified models as a (flowrate into the first cell). In order to develop these models, the effect
complement to an existing dynamic flotation model. Future work is of air flowrate on the pulp level was explicitly ignored by Jämsä-Jounela
therefore needed, and it should focus on the implementation of this et al. (2003). It was also assumed that froth height is negligible in
approach, which could eventually lead to better performance of flota­ comparison to the pulp level. However, it has been demonstrated that air
tion MPC strategies. flowrate directly affects the pulp level in aerated tanks such as those
Another approach that has attracted much attention has been the used in froth flotation (Shean et al., 2018) and, therefore, the effect of
development of models based on hydraulic balances, considering vari­ changes in air flowrate should not be ignored in future studies on
ables that are commonly measured in industrial flotation cells, such as flotation control, especially given the recent focus on optimising per­
the inlet and outlet pulp flowrates. Hydraulic-based models for flotation formance by varying the air flowrate into the cell.
control is the focus of the next section. Similar, but simpler hydraulic balances were proposed by Zaragoza
and Herbst (1989), as shown in Eqs. (68) and (69) for the pulp and froth
3.2.2. Hydraulic models phase, respectively:
Hydraulic models are based on the continuity equation. Jämsä-
dV LP
Jounela et al. (2003) used this principle in order to implement a pulp = QFD − QT − QE + QR , (68)
level control strategy for six flotation cells. One SISO (single-input dt

13
P. Quintanilla et al. Minerals Engineering 162 (2021) 106718

dV LF
= QE − QR − QC , (69)
dt

where QE is the flowrate of entrained water, QR is the flowrate of water

draining back to the pulp, QC is the water leaving the flotation tank with
the concentrate, QT is the tailings flowrate, QFD is the water fed to the
cell, VLP is the volume of the liquid in the pulp and VLF is the volume of
liquid in the froth. Kalman filters were used to implement the on-line
estimation equations of the simplified models (i.e. Eqs. (47), (48),
(68), and (69)) proposed by Zaragoza and Herbst (1989).
A hydraulic approach was also taken by Putz and Cipriano (2015) to
calculate the dynamics of the pulp level (hpj ) for each cell, j. The pulp
level was determined from the hydraulic model developed by Jämsä-
j
Jounela et al. (2003), but with the addition of logic variables δ1 (i.e. Fig. 8. Four states for particles in a flotation cell: (1) Free particles in the pulp
variables that can take a value of 0 or 1, depending on the conditions phase, (2) Attached particles in the pulp phase, (3) Free particles in the froth
previously defined) as: phase and (4) attached particles in the froth phase. (Adapted from Herbst and
Harris (2007), Jovanović et al. (2015)).
dhpj 1( )
= QFj − QTj − QCj δj1 . (70)
dt Aj the pulp phase (Eq. (75)):

where Aj is the cross-sectional area of the cell j, and QFj , QTj and QCj are gd2BP ρp [ ]1.39
vgas = 1 − εg . (75)
the feed, tailings and concentrate flowrates from the cell j, respectively. 18μ
Unlike the model proposed by Jämsä-Jounela et al. (2003), the hy­
The term Xρ is the ratio of the average to surface gas density, which
draulic model presented in Eq. (70) does not include the valve coeffi­
can be calculated as follows:
cient, and rather integrates logical rules as part of the hybrid models. In
all previous hydraulic models presented, it was assumed that the change ρ g(1− εg )hp
(p ) − εg
in the pulp level was the result of only the pulp flowrate into and out of ρp g(1− εg )hp
the cell. However, it has been demonstrated that changes in the gas
P0 ln +1
ρgAve P0

= Xρ ≈ . (76)
flowrate can also have an appreciable impact on the pulp height (Shean ρg0 1 − εg
et al., 2018). These effects are linked to changes in (1) gas hold-up
(Vinnett et al., 2014), and (2) bubble size (Gorain et al., 1995; Nesset For a more in-depth description of the derivation of Eqs. (72) and
et al., 2006). (76), the reader is referred to the Appendix A in Shean et al. (2018). It
The effect of air flowrate on pulp level control was studied by Shean should be noted that Eqs. (73) and (74) are in fact revised versions of
et al. (2018), where a dynamic model was developed and validated in those in Shean et al. (2018) since from Eq. (72) it can be seen that εAve =
both a water-only system and a system with reagents. To carry out the ε0 Xε , rather than εAve = εX0 ε as presented in the original paper. Although it
model development, assumptions such as well-mixed tank, ideal gas law has been demonstrated that the dynamic model in Eq. (74) presents an
for the gas phase, and hydrostatic pressure within the tank, were taken. improvement in predicting pulp level changes, it has never been tested
The change in pulp level was determined using Eq. (71). A relation as part of a level control strategy. In fact, Eq. (74) could be implemented
between average gas hold-up (εave ) and the top of the tank gas hold-up for a better control of the pulp level as it considers not only the effect of
(εg ) was derived, as shown in Eq. (72). the pulp flowrate but also the effect of the air flowrate, specially when
( ) this variable is considered as one that is continuously manipulated.
dV gas d Vsystem εAve d(hεAve ) Nonetheless, it should be noted that only the effect of air flowrate on
= =A , (71)
dt dt dt pulp level was validated at laboratory scale by Shean et al. (2018), and
therefore there is still scope to validate the effect of pulp flowrate (QF
where:
and Qpulpout ).
( )
εAve P
(
ρ g 1 − εg hp
) Whereas hydraulic models, such as those proposed by Jämsä-Jounela
= Xε ≈ ( 0 ) ln p +1 . (72) et al. (2003), Zaragoza and Herbst (1989) and Putz and Cipriano (2015),
εg ρp g 1 − εg hp P0
consider the effect of operational variables (i.e. flowrates) that are
Thus: typically directly measured, other phenomena within the pulp and/or
( ) froth phase that are difficult to measure are ignored, such as attach­
dV gas d hp εg Xε ment/detachment processes, bubble coalescence, and liquid and solid,
=A . (73)
dt dt among others. The attachment/detachment phenomena, however, have
The change in gas hold-up (εg ) for each bubble size class i, can be been included when population balance models are taken into account
determined by Eq. (74), which is a function of air flowrate (QA ), feed for control strategies. In the next section, the population balance
flowrate (QF ), and the total pulp out of the cell (Qpulp,out ), which is the approach is discussed in detail.
sum of tailings and concentrate flowrates:
( ) [ ] [ ]( ) 3.2.3. Population balance models
d Xρ εgi Xε
=
1 QA
− vgas εg,i −
1 QF Qpulp,out
−
εgi Xε Another approach to phenomenological modelling is the use of
dt 1 − εg,total Xε h0 A h0 A A 1 − εg,total Xε population balance models, which can be applied for both simulation
(74) and control purposes (Herbst and Harris, 2007). Mika and Fuerstenau
(1969) described population balances for four states for particles in a
where vgas is the upward gas velocity of the bubble from the pulp phase flotation cell, as shown in Fig. 8, which are: (1) free in the pulp phase,
to the interface, which is a function of gas hold-up and the bubble size in (2) attached to bubbles in the pulp phase, (3) free in the froth phase (i.e.

14
P. Quintanilla et al. Minerals Engineering 162 (2021) 106718

entrained), or (4) attached in the froth phase.

Eq. (77) is the general population balance equation that represents
each of the four possible states (Herbst and Harris, 2007; Herbst and
Flintoff, 2012; Jovanović et al., 2015), for every mineral particle size i,
and mineral particle composition j.

∂V ψ j ∑k
= ψ j,IN QF − ψ j,OUT Qpulpout − kji ψ j V, (77) Fig. 9. Operating conditions in a flotation tank proposed by Putz and Cipriano
∂t
(2015). Mode 1 represents an operation with the presence of both pulp and
i=1

froth phases, without froth overflow; Mode 2 represents a normal operation

where ψ j is the concentration of mineral particles, V is the volume of the
with the presence of both pulp and froth phases, and also froth overflow; and
flotation cell, QF and Qpulpout are the volumetric flowrate into and out of Mode 3 represents an operation with the presence of pulp phase only, with
the cell, respectively, and kji is the particle transfer rate between the overflow of pulp phase at the top of the flotation cell.
states. Each of the four states established in the population balance was
modelled as shown in Eqs. (78)–(81), as follows:

1. Free particles in the pulp phase: Eq. (78) shows the population bal­ d( )
ance for unattached particles in the pulp phase, at time t. This bal­ VLF ψ LF = K FAT LF FDT
ij ij VLF ψ ij +K ij VBF ψ BF LP R LF LF
ij +QE ψ ij − K ij QR ψ ij − QC ψ ij
dt
ance indicates that the accumulation of free particles of size i and (80)
( )
d
mineralogical species j in the pulp phase (dta VLP ψ LP
ij ) is a function

of the rate of attachment (KPAT LP PDT BP

ij VLP ψ ij ) and detachment (Kij VBP ψ ij ) 4. Attached particles in the froth phase: Eq. (81) shows the population
for particles in the pulp, the feed of particles into the flotation cell balance for those particles in the froth that are attached to bubbles.
(QFeed ψ Feed This balance indicates that the accumulation of attached particles in
ij ), the flowrate of particles leaving the cell with the tailings ( )
(QT ψ LP
ij ), the flowrate of particles draining from the froth carried by the froth phase (dtd VBF ψ BF
ij ) is a function of the rate of attachment
water (KRij QR ψ LF
ij ), and the flowrate of entrained particles into the (KFAT LF FDT BF
ij VLF ψ ij ) and detachment (Kij VBF ψ ij ) of particles in the froth
froth (QE ψ LP
ij ). phase, the rate of particles that are carried by air bubbles from the
d( ) pulp phase (QA ψ BP
ij ), and the rate of particles that are attached into
VLP ψ LP = − K PAT LP PDT
ij VLP ψ ij + K ij VBP ψ BP Feed
ij + QFeed ψ ij − QT ψ LP
dt ij ij
bubbles in the froth that reports to the concentrate (QAC ψ BF
ij ).

+K Rij QR ψ LF LP
ij − QE ψ ij
d( )
VBF ψ BF
ij + kFDT
ij VBF ψ BF FAT LF BP BF
ij − kij VLF ψ ij = QA ψ ij − QAC ψ ij
(81)
(78) dt

2. Attached particles in the pulp phase: Eq. (79) shows the population
balance for particles that are attached to a bubble in the pulp phase, Parameters, such as rate constants, water entrainment/drainage and
at time t. This balance indicates that the accumulation of particles water into the concentrate, can be determined empirically as shown in
( ) Section 3.1, in Eqs. (23)–(33). In the study carried out by Bascur (1982),
attached to bubbles in the pulp phase (dtd VLP ψ BP
ij ) is a function of phenomenological models were developed with a focus on automatic
control, by considering three types of particles: free valuable minerals,
the rate of attachment (KPAT LP PDT
ij VLP ψ ij ) and detachment (Kij VBP ) of free gangue, and locked particles.
particles, the transport of particles attached to bubbles to the froth The aforementioned population balance models are used to represent
(QA ψ BP
ij ), and the particles attached to bubbles leaving the cell the attachment and detachment phenomena, as well as the solids
through tailings (QAT ψ BP
AT ). transfer between the pulp phase and froth phase. However, these models
do not fully represent the phenomena that occur in a flotation cell,
d( )
VLP ψ BP
ij = K PAT LP PDT
ij VLP ψ ij − K ij VBP ψ BP BP LP
ij + QA ψ ij − QAT ψ ij
(79) specially in the froth phase. To date, physics-based models for the froth
dt
phase have not been developed for control purposes. While phenome­
nological models based on kinetics and population balances work well to
3. Free particles in the froth phase: Eq. (80) shows the population represent the pulp phase, they do not adequately represent the froth
balance for particles entrained with the liquid in the froth phase. This phase, with the assumption of a well-mixed system being particularly
balance indicates that the accumulation of free particles in the froth problematic. They also do not incorporate the important stability effects,
)
( such as bubble coalescence and bursting which are important drivers in
phase (dtd VLF ψ LF
ij ) is a function of the rate of attachment the overall performance of froths.
(KFAT LF FDT BF
ij VLF ψ ij ) and detachment (Kij VBF ψ ij ) of particles in the froth
phase, the rate of particles carried due to water entrainment from the 3.3. Hybrid models
pulp (QE ψ LP
ij ), the rate of particles carried out of the froth due to
water drainage (KRij QR ψ LF Hybrid models are used to represent a process by using both
ij ), and the rate of non-attached particles in
continuous and discrete variables. The discrete variables are logic rules
the froth that overflows into the concentrate launder (QC ψ LF
ij ).
that represent different operating conditions of the process (also known
as “modes”) that generate changes in the continuous variables. The first

15
P. Quintanilla et al. Minerals Engineering 162 (2021) 106718

hybrid model predictive control for flotation, based on PWARX (Piece­ process.
Wise AutoRegresive eXogenous) models (Heemels et al., 2001), was
developed by Putz and Cipriano (2015). The PWARX system was used to 4. Conclusions
approximate non-linear process dynamics, such as the froth flotation
process, by using linearised models at different operating conditions Numerous studies have established that the efficiency of flotation
(Sontag, 1981). control can be improved by using advanced controllers, such as Model
In order to implement the hybrid MPC, Putz and Cipriano (2015) Predictive Control (MPC). MPC is a powerful control strategy for com­
defined three modes, as shown in Fig. 9, as follows: (1) flotation cell with plex processes such as froth flotation. This control strategy optimises the
the presence of pulp and froth phases, but no concentrate overflow, (2) process by using explicit models able to predict its outputs, minimising a
normal operation (presence of the pulp and froth phases, as well as cost function that depends on process variables and process constraints.
concentrate overflow), and (3) operation with overflow of the pulp The most crucial part of this control strategy is the model development
phase (i.e. no presence of the froth phase). In order to simplify the itself, specially in a complex process such as froth flotation, due to its
formulation, only the first two operating conditions were taken into dynamic nature and the fact it can be affected by a great number of
account, with the third one considered as an operating (hip ⩽himax , i.e. variables.
pulp height cannot be greater than the cell height). This review has, for the first time in the literature, classified and
In order to implement the hybrid MPC, assumptions were made such analysed models used for flotation MPC strategies, providing a frame­
as perfect mixing, constant air flowrate, and constant tank cross- work for future studies. The models used for MPC strategies were clas­
sectional area, and considering two mineralogical classes (high and sified as empirical, phenomenological or hybrid models. Empirical
low chalcopyrite grade). This last consideration on mineralogical classes models have been mainly developed to determine flotation rate con­
is, in fact, not ideal as the mineralogy have a significant impact on the stants, drainage rate constant, concentrate volumetric flowrate, among
concentrate economic value. More realistic results could have been other parameters, by using fits to plant data. Although these models are
obtained if a distributed mineralogy, for example, was taken into ac­ cost-effective, they can only be used when the process is under a well-
count instead. known range of operating conditions and, therefore, their robustness
j j outside that range is highly debatable. Phenomenological models, as
Two logical variables were defined as δ1 and δ2 . Depending on the
well as hybrid models, can be applied to into any flotation system since
operating conditions, the logical variables can take the value of 0 or 1.
they are derived from first principles. In this review, phenomenological
Eq. (82) and Eq. (83) show the logical rules that were defined to
models were classified into three types: kinetic, hydraulic and popula­
implement the hybrid MPC by Putz and Cipriano (2015).
tion balance models.
⎧
⎨ 1 if hj ⩽hj The analysis of the literature on the topic has shown that little evi­
(82) dence on successful MPC implementation in flotation at industrial scale
j p max
δ1 = ,
⎩ 0 if hjp > hjmax
is available. There remains a need for further research to enhance
⎧ modelling for flotation control purposes. To date, MPC studies have
⎨1 if hjp + hif > hjmax tended to use kinetic models to phenomenologically represent the
δj2 = (83) attachment-detachment of mineral particles in the pulp phase and the
⎩0 if hjp + hjf ⩽hjmax
froth phase. While kinetic models can accurately represent the phe­
nomena in the pulp phase, these models are not suitable for the froth
j
where hjp and hf are the pulp height and froth depth of the cell j, phase, as more complex phenomena are involved. In fact, froth phase
j j
respectively. Both logical variables δ1 and δ2 , can be grouped in a vector phenomena such as bubble coalescence, bubble bursting, and liquid and
[ ] solid motion, are key drivers of froth performance. There is therefore an
that contains each value, as δj1 δj2 . The three operating conditions
important opportunity to develop suitable models that take into account
previously defined and their respective logical states are presented in the physics of the froth phase and can be implemented into MPC stra­
Fig. 9. The use of only logical rules, however, was not enough to effec­ tegies. Addressing the gap between models for the pulp phase and for the
tively implement a model predictive controller. For this reason, empir­ froth phase suitable to be implemented into MPC strategies will pave the
ical (Eqs. (6), (7), (15), and (16)) phenomenological models (kinetic way for improving the overall performance of the flotation process.
balances: Eqs. (45) and (46); hydraulic model: Eq. (70)) were also
implemented by Putz and Cipriano (2015). Declaration of Competing Interest
Simulations were run to test the proposed hybrid flotation models,
with the feed flowrate and the feed grade as the measured disturbance The authors declare that they have no known competing financial
variables. The controlled variable was the pulp height and the final interests or personal relationships that could have appeared to influence
tailings grade, which was manipulated by controlling the position of the the work reported in this paper.
tailings valve. The effect that air flowrate and regent addition rate were
thus ignored in terms of their influence on the pulp level. The control
Acknowledgements
aim was to minimise the error between a reference and the final tail
grade, while minimising the variations in the manipulated variables.
Paulina Quintanilla would like to acknowledge the National Agency
Although other hybrid models have been developed for analysis or
for Research and Development (ANID) for funding this research with a
design purposes, such as Gupta et al. (1999), to the knowledge of the
scholarship from “Becas Chile”. Dr Diego Mesa is also acknowledged for
authors, the hybrid model implemented into a control strategy by Putz
valuable comments on this manuscript.
and Cipriano (2015) is the only one of its kind and, thus, no comparison
can be made with other control studies. It should, however, be noted
Appendix A
that is important to identify gaps in knowledge that can lead to further
research and, hence, to find a way to optimally control the flotation
Table A1.

16
 P. Quintanilla et al.
Table A1
Control variables most commonly used in MPC for flotation.
Study Input variables Output variables

Manipulated Disturbances Controlled Internal states

Zaragoza and Herbst Set points of the regulatory control Flotation feed disturbances in the first cells of the circuit (using Kalman Pulp level Mass of chalcopyrite in the pulp and froth phases
(1989) loops filter) Mass of locked particles in the pulp and froth phases
Mass of gangue in the pulp and froth phases
Mass of free pyrite in the pulp and froth phases
Volume of liquid in pulp and froth phases (in cell 1 and
2)
Perez-Correa et al. (1998) Collector doses Copper head grade Concentrate grade Not reported explicitly
Frother addition rate Feed flowrate Tailings grade
Pulp level set point Average feed particle size Metallurgical recovery
Iron feed content
Feed pH
Maldonado et al. (2007) Output control valve of tailings Not reported explicitly Froth depth Not reported explicitly
17

flowrate Bias rate

Wash-water flowrate set point Gas hold-up
Air flowrate set point
Maldonado et al. (2007) Output control valve of tailings Not reported explicitly Pulp level at the bank j Tailings grade
flowrate from each bank Cumulative concentrate solid mass flowrate
Grade in the concentrate
Maldonado et al. (2009) Output control valve of tailings Not reported explicitly Froth depth Not reported explicitly
flowrate Bias rate
Wash water flowrate set point Collection zone gas hold-up
Air flowrate set point
Maldonado et al. (2010) Superficial water velocity set point Changes on superficial gas velocity Sauter mean diameter Not reported explicitly
Frother concentration
Putz and Cipriano (2015) Output control valve of tailings Feed flowrate Pulp level Not reported explicitly
flowrate from each bank Copper head grade Final tailings grade
Tian et al. (2018) Feed velocity set point Not reported explicitly Mass concentration of solid Not reported explicitly
particles
in the air phase of froth overflow

Minerals Engineering 162 (2021) 106718

P. Quintanilla et al. Minerals Engineering 162 (2021) 106718

References disturbance rejection dynamics. IEE Proc.-Control Theory Appl. 147, 465–475.
https://doi.org/10.1049/ip-cta:20000443.
Dinariev, O.Y., Evseev, N.V., 2018. Modelling of flotation processes by density functional
Al-Thyabat, S., 2008. On the optimization of froth flotation by the use of an artificial
hydrodynamics. Miner. Eng. https://doi.org/10.1016/j.mineng.2018.06.013.
neural network. J. China Univ. Min. Technol. 18, 418–426. https://doi.org/
Faulkner, B., 1966. Computer control improves metallurgy at Tennessee Copper’s
10.1016/S1006-1266(08)60087-5.
flotation plant. Min. Eng. 18, 53–57.
Aldrich, C., Marais, C., Shean, B., Cilliers, J., 2010. Online monitoring and control of
Ferreira, J.P., Loveday, B.K., 2000. Improved model for simulation of flotation circuits.
froth flotation systems with machine vision: A review. Int. J. Miner. Process. 96,
Miner. Eng. https://doi.org/10.1016/S0892-6875(00)00129-1.
1–13. https://doi.org/10.1016/J.MINPRO.2010.04.005.
Finch, J.A., Dobby, G.S., 1991. Column flotation: A selected review. Part I. Int. J. Miner.
Alves dos Santos, N., Savassi, O., Peres, A.E.C., Martins, A.H., 2014. Modelling flotation
Process. https://doi.org/10.1016/0301-7516(91)90062-N.
with a flexible approach – Integrating different models to the compartment model.
Forbes, G., 2007. Texture and bubble size measurements for modelling concentrate grade
Miner. Eng. 66–68, 68–76. https://doi.org/10.1016/J.MINENG.2014.05.007.
in flotation froth systems (Phd thesis). University of Cape Town.
https://www.sciencedirect.com/science/article/pii/S0892687514001769.
Fu, Y., Aldrich, C., 2019. Flotation froth image recognition with convolutional neural
Arbiter, N., Harris, C.C., 1962. Flotation kinetics. In: Fuer, D.W. (Ed.), Froth Flotation.
networks. Miner. Eng. 132, 183–190. https://doi.org/10.1016/J.
Edwards Brothers Inc, New York, pp. 215–246.
MINENG.2018.12.011.
Bartolacci, G., Ourriban, M., Bouajila, A., Gomez, C., Finch, J., Goyette, G., 2008. On-line
Garcia-Zuniga, H., 1935. The efficiency obtained by flotation is an exponential funcion of
use of hydrodynamic sensors to improve metallurgical performance of flotation
time. Soc. Nac. Minera 47, 83–86.
machine. In: Gomez, C., Kuyvenhoven, R., Casali, A. (Eds.), V International Mineral
Gharai, M., Venugopal, R., 2016. Modeling of flotation process - An overview of different
Processing Seminar. PROCEMIN, Santiago, Chile, pp. 357–366.
approaches. Miner. Process. Extr. Metall. Rev. 37, 120–133. https://doi.org/
Bartolacci, G., Pelletier, P., Tessier, J., Duchesne, C., Bossé, P.A., Fournier, J., 2006.
10.1080/08827508.2015.1115991.
Application of numerical image analysis to process diagnosis and physical parameter
Gomez, C.O., Cortés-López, F., Finch, J.A., 2003. Industrial testing of a gas holdup sensor
measurement in mineral processes–Part I: Flotation control based on froth textural
for flotation systems. Miner. Eng. https://doi.org/10.1016/S0892-6875(03)00083-9.
characteristics. Miner. Eng. 19, 734–747. https://doi.org/10.1016/J.
Gorain, B.K., Franzidis, J.P., Manlapig, E.V., 1995. Studies on impeller type, impeller
MINENG.2005.09.041.
speed and air flow rate in an industrial scale flotation cell - Part 1: Effect on bubble
Bascur, O.A., 1982. Modelling and computer control of a flotation cell. University of
size distribution. Miner. Eng. https://doi.org/10.1016/0892-6875(95)00025-L.
Utah, Salt Lake City.
Gupta, S., Liu, P.H., Svoronos, S.A., Sharma, R., Abdel-Khalek, N.A., Cheng, Y., El-
Bergh, L., Yianatos, J., 1994. Experimental studies on flotation column dynamics. Miner.
Shall, H., 1999. Hybrid first-principles/neural networks model for column flotation.
Eng. 7, 345–355. https://doi.org/10.1016/0892-6875(94)90075-2.
AIChE J. 45, 557–566. https://doi.org/10.1002/aic.690450312.
Bergh, L., Yianatos, J., 1996. Hierarchical control strategy in columns at El Teniente. In:
Gupta A., Yan D.S., 2006. Process Control. Mineral Processing Design and Operation, pp.
Proceedings of the International Symposium on Column Flotation, Montreal,
622–671. doi: 10.1016/B978-044451636-7/50019-X (Chapter 18).
Canada, pp. 369–380.
Hadler, K., Cilliers, J.J., 2009. The relationship between the peak in air recovery and
Bergh, L., Yianatos, J., Pino, C., 2013. Advances in developing supervisory control
flotation bank performance. Miner. Eng. https://doi.org/10.1016/j.
strategies for flotation plants. IFAC Proc. Vol. 46, 110–115. https://doi.org/
mineng.2008.12.004.
10.3182/20130825-4-US-2038.00003.
Hadler, K., Smith, C.D., Cilliers, J.J., 2010. Recovery vs. mass pull: The link to air
Bergh, L.G., 2016. Artificial intelligence in mineral processing plants: an overview. In:
recovery. Miner. Eng. 23, 994–1002. https://doi.org/10.1016/j.
International Conference on Artificial Intelligence: Technologies and Applications
mineng.2010.04.007.
(ICAITA 2016), pp. 278–281.
Heemels, W.P.M.H., De Schutter, B., Bemporad, A., 2001. Equivalence of hybrid
Bergh, L.G., Yianatos, J.B., 2003. Flotation column automation: State of the art. Control
dynamical models. Automatica. https://doi.org/10.1016/S0005-1098(01)00059-0.
Eng. Pract. https://doi.org/10.1016/S0967-0661(02)00093-X.
Herbst, J.A., Flintoff, B., 2012. Recent advances in modeling, simulation, and control of
Bergh, L.G., Yianatos, J.B., 2011. The long way toward multivariate predictive control of
mineral processing operations. In: Separation Technologies for Minerals, Coal, and
flotation processes. J. Process Control 21, 226–234. https://doi.org/10.1016/j.
Earth Resources, pp. 667–680.
jprocont.2010.11.001.
Herbst, J.A., Harris, M., 2007. Modeling and Simulation of Industrial Flotation Processes.
Bergh, L.G., Yianatos, J.B., Acuña, C.A., Pérez, H., López, F., 1999. Supervisory control at
In: Froth Flotation: A Century of Innovation, pp. 757–777.
Salvador flotation columns. Miner. Eng. 12, 733–744. https://doi.org/10.1016/
Herbst, J.A., Pate, W.T., Oblad, A.E., 1992. Model-based control of mineral processing
S0892-6875(99)00060-6.
operations. Powder Technol. 69, 21–32.
Bergh, L.G., Yianatos, J.B., Leiva, C.A., 1998. Fuzzy supervisory control of flotation
Hodouin, D., 2011. Methods for automatic control, observation, and optimization in
columns. Miner. Eng. https://doi.org/10.1016/s0892-6875(98)00059-4.
mineral processing plants. J. Process Control 21, 211–225. https://doi.org/10.1016/
Bertsekas, D.P., 1995. Dynamic Programming and Optimal Control. Athena Scientific.
j.jprocont.2010.10.016.
Bonifazi, G., Serranti, S., Volpe, F., Zuco, R., 2001. Characterisation of flotation froth
Hodouin, D., Jämsä-Jounela, S.L., Carvalho, M.T., Bergh, L., 2001. State of the art and
colour and structure by machine vision. Comput. Geosci. 27, 1111–1117. https://
challenges in mineral processing control. Control Eng. Practice 9, 995–1005. https://
doi.org/10.1016/S0098-3004(00)00152-7.
doi.org/10.1016/S0967-0661(01)00088-0.
Bordons, C., 2000. Control predictivo: metodología, tecnología y nuevas perspectivas.
Holtham, P., Nguyen, K., 2002. On-line analysis of froth surface in coal and mineral
Technical Report. Universidad de Sevilla, Aguadulce, Almería.
flotation using JKFrothCam. Int. J. Miner. Process. 64, 163–180. https://doi.org/
Bouchard, J., Desbiens, A., del Villar, R., Nunez, E., 2009. Column flotation simulation
10.1016/S0301-7516(01)00070-9.
and control: An overview. Miner. Eng. 22, 519–529. https://doi.org/10.1016/j.
Hsu, K.l., Gupta, H.V., Sorooshian, S., 1995. Artificial Neural Network Modeling of the
mineng.2009.02.004.
Rainfall-Runoff Process. Water Resources Research doi:10.1029/95WR01955.
Camacho, E.F., Bordons, C., 2007. Model Predictive Control. https://doi.org/10.1007/
Humaloja, J.P., Dubljevic, S., 2018. Linear Model Predictive Control for Schrödinger
978-0-85729-398-5 arXiv:arXiv:1011.1669v3.
Equation. Proceedings of the American Control Conference 2018-June, 2569–2574.
Carr, D., Dixon, A., Tiili, O., 2009. Optimising Large Flotation Cell Performance Through
doi:10.23919/ACC.2018.8431686.
Advanced Instrumentation and Control. In: Tenth Mill Operator’s Conference,
Hyötyniemi, H., Ylinen, R., 2000. Modeling of visual flotation froth data. Control Eng.
Adelaide, SA, pp. 299–304.
Practice 8, 313–318. https://doi.org/10.1016/S0967-0661(99)00187-2.
Casali, A., Gonzalez, G., Agusto, H., Vallebuona, G., 2002. Dynamic simulator of a
Jahedsaravani, A., Marhaban, M.H., Massinaei, M., Saripan, M.I., Noor, S.B., 2016.
rougher flotation circuit for a copper sulphide ore. Miner. Eng. 15, 253–262. https://
Froth-based modeling and control of a batch flotation process. Int. J. Miner. Process.
doi.org/10.1016/S0892-6875(02)00016-X.
146, 90–96. https://doi.org/10.1016/j.minpro.2015.12.002.
Chelgani, S.C., Shahbazi, B., Rezai, B., 2010. Estimation of froth flotation recovery and
Jämsä-Jounela, S.L., Dietrich, M., Halmevaara, K., Tiili, O., 2003. Control of pulp levels
collision probability based on operational parameters using an artificial neural
in flotation cells. Control Eng. Practice 11, 73–81. https://doi.org/10.1016/S0967-
network. Int. J. Min. Metall. Mater. 17, 526–534. https://doi.org/10.1007/s12613-
0661(02)00142-9.
010-0353-1.
Jovanović, I., Miljanović, I., 2015. Contemporary advanced control techniques for
Cipriano, A., Guarini, M., Vidal, R., Soto, A., Sepúlveda, C., Mery, D., Briseño, H., 1998.
flotation plants with mechanical flotation cells - A review. Miner. Eng. 70, 228–249.
A real time visual sensor for supervision of flotation cells. Miner. Eng. https://doi.
https://doi.org/10.1016/j.mineng.2014.09.022.
org/10.1016/s0892-6875(98)00031-4.
Jovanović, I., Miljanović, I., Jovanović, T., 2015. Soft computing-based modeling of
Citir, C., Aktas, Z., Berber, R., 2003. Off-line image analysis for froth flotation of coal.
flotation processes – A review. Miner. Eng. 84, 34–63. https://doi.org/10.1016/J.
Comput. Aided Chem. Eng. 14, 605–610. https://doi.org/10.1016/S1570-7946(03)
MINENG.2015.09.020.
80182-7.
Kelsall, D., 1961. Application of probability in the assessment of flotation systems. Bull.
Cubillos, F.A., Lima, E.L., 1997. Identification and optimizing control of a rougher
Inst. Min. Metall. 650, 191–204.
flotation circuit using an adaptable hybrid-neural model. Miner. Eng. 10, 707–721.
Kracht, W., Gomez, C.O., Finch, J.A., 2008. Controlling bubble size using a frit and sleeve
https://doi.org/10.1016/s0892-6875(97)00050-2.
sparger. Miner. Eng. 21, 660–663. https://doi.org/10.1016/j.mineng.2007.12.009.
Del Villar, R., Grégoire, M., Pomerleau, A., 1999. Automatic control of a laboratory
Laurila, H., Karesvuori, J., Tiili, O., 2002. Strategies for Instrumentation and Control of
flotation column. Miner. Eng. 12, 291–308. https://doi.org/10.1016/s0892-6875
Flotation Circuits. In: Mineral Processing Plant Design, Practice and Control
(99)00007-2.
Proceedings, pp. 2174–2195.
Desbiens, A., Hodouin, D., Mailloux, M., 1998. Nonlinear predictive control of a rougher
Leonard, R., Lemlich, R., 1965. Laminar longitudinal flow between close-packed
flotation unit using local models. IFAC Proc. Vol. 31, 287–292. https://doi.org/
cylinders. J. Chem. Eng. Sci. 20, 790–791.
10.1016/S1474-6670(17)35893-7.
Lewis, H., 1997. The foundations of fuzzy control. Plenum Press, New York.
Desbiens, A., Hodouin, D., Najim, K.I., Flament, F., Universitaire, C., Gk, C., 1994. Long-
Liao, Y., Liu, J., Wang, Y., Cao, Y., 2011. Simulating a fuzzy level controller for flotation
range predictive control of a rougher flotation unit. Miner. Eng. 7, 21–37.
columns. Min. Sci. Technol. https://doi.org/10.1016/j.mstc.2011.05.038.
Desbiens, A., Hodouin, D., Plamondon, E., 2000. Global predictive control: A unified
control structure for decoupling setpoint tracking, feedforward compensation and

18
P. Quintanilla et al. Minerals Engineering 162 (2021) 106718

Liu, J., MacGregor, J., Duchesne, C., Bartolacci, G., 2005. Flotation froth monitoring Nguyen, A.V., Ralston, J., Schulze, H.J., 1998. On modelling of bubble-particle
using multiresolutional multivariate image analysis. Miner. Eng. 18, 65–76. https:// attachment probability in flotation. Int. J. Miner. Process. 53, 225–249. https://doi.
doi.org/10.1016/J.MINENG.2004.05.010. org/10.1016/S0301-7516(97)00073-2.
Liu, J.J., MacGregor, J.F., 2008. Froth-based modeling and control of flotation processes. Niemi, A.J., Ylinen, R., Hyötyniemi, H., 1997. On characterization of pulp and froth in
Miner. Eng. 21, 642–651. https://doi.org/10.1016/j.mineng.2007.12.011. cells of flotation plant. Int. J. Miner. Process. 51, 51–65. https://doi.org/10.1016/
Lynch, A., Johnson, N., Manlapig, E., Thorne, C., 1981. Mineral and Caol Flotation s0301-7516(97)00042-2.
Circuits - Their Simulation and Control. Netherlands. Oosthuizen, D.J., Craig, I.K., 2019. Predicting optimal operating points by modelling
Maciejovski, J., 2002. Predictive control with constraints. Prentice Hall. different flotation mechanisms. IFAC-PapersOnLine 52, 60–65. https://doi.org/
Maldonado, M., 2010. Advances in Estimation and Control for Flotation Columns (Ph.D. 10.1016/j.ifacol.2019.09.164.
thesis). Université Laval. Oosthuizen, D.J., Craig, I.K., Jämsä-Jounela, S.L., Sun, B., 2017. On the current state of
Maldonado, M., Desbiens, A., Del Villar, R., 2009. Potential use of model predictive flotation modelling for process control. IFAC-PapersOnLine 50, 19–24. https://doi.
control for optimizing the column flotation process. Int. J. Miner. Process. 93, 26–33. org/10.1016/j.ifacol.2017.12.004.
https://doi.org/10.1016/j.minpro.2009.05.004. Osorio, D., Pérez-Correa, J.R., Cipriano, A., 1999. Assessment of expert fuzzy controllers
Maldonado, M., Desbiens, A., Del Villar, R., Girgin, E., Gomez, C., 2008. On-line for conventional flotation plants. Miner. Eng. https://doi.org/10.1016/S0892-6875
estimation of bubble size distributions using Gaussian mixture models. In: (99)00120-X.
Kuyvenhoven, C., Casali, A. (Eds.), V International Mineral Processing Seminar. Perez-Correa, R., Gonzalez, G., Casali, A., Cipriano, A., Barrera, R., Zavala, E., 1998.
PROCEMIN, Santiago, Chile, pp. 389–398. Dynamic modelling and advanced multivariable control of conventional flotation
Maldonado, M., Desbiens, A., Del Villar, R., Poulin, E., Riquelme, A., 2010. Nonlinear circuits. Miner. Eng. 11, 333–346. https://doi.org/10.1016/s0892-6875(98)00012-0.
control of bubble size in a laboratory otation column, vol. 43. Cape Town. doi: Pitt, J., 1968. The development of systems for continuous optimal control of flotation
10.3182/20100802-3-ZA-2014.00004. plants by computer. In: System Dynamics and Automatic Control in Basic Industries
Maldonado, M., Desbiens, A., Del Villar, R., Quispe, R., 2007. Towards the optimization (IFAC Symposium), Sydney, pp. 165–171.
of flotation columns using predictive control. IFAC Proc. Vol. (IFAC-PapersOnline) Polat, M., Chander, S., 2000. First-order flotation kinetics models and methods for
12, 75–80. https://doi.org/10.3182/20070821-3-CA-2919.00011. estimation of the true distribution of flotation rate constants. Int. J. Miner. Process.
Maldonado, M., Sbarbaro, D., Lizama, E., 2007. Optimal control of a rougher flotation 58, 145–166. https://doi.org/10.1016/S0301-7516(99)00069-1.
process based on dynamic programming. Miner. Eng. 20, 221–232. https://doi.org/ Prakash, R., Majumder, S.K., Singh, A., 2018. Flotation technique: Its mechanisms and
10.1016/j.mineng.2006.08.015. design parameters. Chem. Eng. Process. Process Intensification 127, 249–270.
Massinaei, M., Doostmohammadi, R., 2010. Modeling of bubble surface area flux in an https://doi.org/10.1016/j.cep.2018.03.029.
industrial rougher column using artificial neural network and statistical techniques. Putz, E., Cipriano, A., 2015. Hybrid model predictive control for flotation plants. Miner.
Miner. Eng. https://doi.org/10.1016/j.mineng.2009.10.005. Eng. 70, 26–35. https://doi.org/10.1016/j.mineng.2014.08.013.
Mckay, J., Inchausti, R., 1996. Expert supervisory control of flotation columns. In: Qin, S., Badgwell, T.A., 1997. An overview of industrial model predictive control
Proceedings of the International Symposium on Column Flotation, Montreal, technology. Control Eng. Practice doi:10.1.1.52.8909.
Canada, pp. 353–367. Riquelme, A., Desbiens, A., Del Villar, R., Maldonado, M., 2016. Predictive control of the
Mesa, D., Brito-Parada, P.R., 2019. Scale-up in froth flotation: A state-of-the-art review. bubble size distribution in a two-phase pilot flotation column. Miner. Eng. 89,
Sep. Purif. Technol. 210, 950–962. https://doi.org/10.1016/j.seppur.2018.08.076. 71–76. https://doi.org/10.1016/j.mineng.2016.01.014.
Mika, T.S., Fuerstenau, D.W., 1969. A microscopic model of the flotation process. In: Rojas, D., Cipriano, A., 2011. Model Based Predictive Control of a Rougher Flotation
Proceedings of the VIII International Mineral Processing Congress, Leningrad. pp. Circuit Considering Grade Estimation in Intermediate Cells. Technical Report.
246–269. Universidad Nacional de Colombia, Medellin. http://www.redalyc.org/articulo.oa?
Mohanty, S., 2009. Artificial neural network based system identification and model id=49622365004%0AHow.
predictive control of a flotation column. J. Process Control 19, 991–999. https://doi. Rossiter, J., 2003. Model Based Predictive Control: A Practical Approach. Control Series.
org/10.1016/j.jprocont.2009.01.001. Sbarbaro, D., del Villar, R., 2010. Advanced control and supervision for mineral
Moolman, D., Aldrich, C., Van Deventer, J., 1995a. The analysis of videographic data processing. Adv. Ind. Control. https://doi.org/10.1007/978-1-84996-106-6.
with neural nets. Chim. Acta Slovenica 42, 137–142. Schubert, H., 2008. On the optimization of hydrodynamics in fine particle flotation.
Moolman, D., Aldrich, C., Van Deventer, J., Stange, W., 1994. Digital image processing as Miner. Eng. 21, 930–936. https://doi.org/10.1016/j.mineng.2008.02.012.
a tool for on-line monitoring of froth in flotation plants. Min. Eng. 17, 1149–1164. Shean, B., Hadler, K., Cilliers, J.J., 2017. A flotation control system to optimise
https://doi.org/10.1016/0892-6875(94)00058-1. performance using peak air recovery. Chem. Eng. Res. Des. 117, 57–65. https://doi.
Moolman, D.W., Aldrich, C., van Deventer, J.S., Stange, W.W., 1995b. The classification org/10.1016/j.cherd.2016.10.021.
of froth structures in a copper flotation plant by means of a neural net. Int. J. Miner. Shean, B., Hadler, K., Neethling, S., Cilliers, J.J., 2018. A dynamic model for level
Process. 43, 193–208. https://doi.org/10.1016/0301-7516(95)00003-V. prediction in aerated tanks. Miner. Eng. 125, 140–149. https://doi.org/10.1016/j.
Moolman, D.W., Aldrich, C., Schmitz, G.P., Van Deventer, J.S., 1996. The mineng.2018.05.030.
interrelationship between surface froth characteristics and industrial flotation Shean, B.J., Cilliers, J.J., 2011. A review of froth flotation control. Int. J. Miner. Process.
performance. Miner. Eng. 9, 837–854. https://doi.org/10.1016/0892-6875(96) 100, 57–71. https://doi.org/10.1016/j.minpro.2011.05.002.
00076-3. Smith, H., Lewis, C., 1969. Computer control experiments at Lake Dufault. Can. I.M.M.
Moolman, D.W., Aldrich, C., Van Deventer, J.S., 1995c. The monitoring of froth surfaces Bulletin 62, 109–115.
on industrial flotation plants using connectionist image processing techniques. Sontag, E., 1981. Nonlinear regulations: the piecewise linear approach. IEEE Trans.
Miner. Eng. 8, 23–30. https://doi.org/10.1016/0892-6875(94)00099-X. Automatic Control 26, 346–358. https://doi.org/10.1109/TAC.1981.1102596.
Moolman, D.W., Aldrich, C., Van Deventer, J.S., Bradshaw, D.J., 1995d. The Tian, Y., Luan, X., Liu, F., Dubljevic, S., 2018. Model predictive control of mineral
interpretation of flotation froth surfaces by using digital image analysis and neural column flotation process. Mathematics 6, 100. https://doi.org/10.3390/
networks. Chem. Eng. Sci. 50, 3501–3513. https://doi.org/10.1016/0009-2509(95) math6060100.
00190-G. Varbanov, R., Forssberg, E., Hallin, M., 1993. On the modelling of the flotation process.
Mular, A.L., 1972. Empirical modelling and optimization of mineral processes. Min. Sci. Int. J. Miner. Process. 37, 27–43. https://doi.org/10.1016/0301-7516(93)90003-S.
Eng. Ventura-Medina, E., Cilliers, J.J., 2000. Calculation of the specific surface area in
Neethling, S., Brito-Parada, P., Hadler, K., Cilliers, J., 2019. The transition from first to flotation. Miner. Eng. 13, 265–275. https://doi.org/10.1016/S0892-6875(00)
zero order flotation kinetics and its implications for the efficiency of large flotation 00006-6.
cells. Miner. Eng. 132, 149–161. https://doi.org/10.1016/J.MINENG.2018.11.039. Verbist, G., Weaire, D., Kraynik, A.M., 1996. The foam drainage equation. J. Phys.
Neethling, S., Cilliers, J., 2009. The entrainment factor in froth flotation: Model for Condens. Matter 8, 3715–3731. https://doi.org/10.1088/0953-8984/8/21/002.
particle size and other operating parameter effects. Int. J. Miner. Process. 93, Veselý, V., Rosinová, D., 2010. Robust Model Predictive Control Design. In: Zheng, T.
141–148. https://doi.org/10.1016/J.MINPRO.2009.07.004. (Ed.), Model Predictive Control, pp. 217–248. doi:10.5772/256.
Neethling, S., Lee, H., Cilliers, J., 2002. A foam drainage equation generalised for all Vinnett, L., Yianatos, J., Alvarez, M., 2014. Gas dispersion measurements in mechanical
liquid contents. J. Phys. Condens. Matter 14, 331–342. flotation cells: Industrial experience in Chilean concentrators. Miner. Eng. https://
Neethling, S.J., Brito-Parada, P.R., 2018. Predicting flotation behaviour – The interaction doi.org/10.1016/j.mineng.2013.12.006.
between froth stability and performance. Miner. Eng. 120, 60–65. https://doi.org/ Wang, G., Ge, L., Mitra, S., Evans, G.M., Joshi, J.B., Chen, S., 2018. A review of CFD
10.1016/j.mineng.2018.02.002. modelling studies on the flotation process. doi:10.1016/j.mineng.2018.08.019.
Neethling, S.J., Cilliers, J.J., 2002. Solids motion in flowing froths. Chem. Eng. Sci. 57, Wang, L., Peng, Y., Runge, K., Bradshaw, D., 2015. A review of entrainment:
607–615. Mechanisms, contributing factors and modelling in flotation. doi:10.1016/j.
Neethling, S.J., Cilliers, J.J., 2008. Predicting air recovery in flotation cells. Miner. Eng. mineng.2014.09.003.
21, 937–943. https://doi.org/10.1016/j.mineng.2008.03.011. Wang, W., Bergholm, F., Yang, B., 2003. Froth delineation based on image classification.
Neethling, S.J., Cilliers, J.J., Woodburn, E.T., 2000. Prediction of the water distribution Miner. Eng. 16, 1183–1192. https://doi.org/10.1016/j.mineng.2003.07.014.
in a flowing foam. Chem. Eng. Sci. 55, 4021–4028. Whiten, W., 1972. Simulation and model building for mineral processing (Ph.D. thesis).
Neethling, S.J., Lee, H.T., Cilliers, J.J., 2003. Simple relationships for predicting the University of Queensland.
recovery of liquid from flowing foams and froths. Miner. Eng. 16, 1123–1130. Yianatos, J.B., Bergh, L.G., 1995. Troubleshooting industrial flotation columns. Miner.
https://doi.org/10.1016/j.mineng.2003.06.014. Eng. https://doi.org/10.1016/0892-6875(95)00121-2.
Nesset, J.E., Hernandez-Aguilar, J.R., Acuna, C., Gomez, C.O., Finch, J.A., 2006. Some Yoon, R.H., Mao, L.Q., 1996. Application of extended DLVO theory.4. Derivation of
gas dispersion characteristics of mechanical flotation machines. Miner. Eng. flotation rate equation from first principles. J. Colloid Interface Sci. 181, 613–626.
807–815. https://doi.org/10.1016/j.mineng.2005.09.045. https://doi.org/10.1006/jcis.1996.0419.
Zaragoza, R., Herbst, J.A., 1989. Model-based feedforward control scheme for flotation
plants. Min. Metall. Process. 177–185.

19