894 XXXI International Mineral Processing Congress 2024 Proceedings/Washington, DC/Sep 29–Oct 3
dependent on process variables such as feed grade, mineral-
ogy, pulp density, reagent type addition rate, froth depth and
air flow rate, displaying complex interactions (Amankwaa-
Kyeremeh et al., 2021a Dankwah et al., 2022a, b Forson
et al., 2022 Mathe et al., 2000). The process complexity
has warranted froth flotation modelling to predict recovery
and grade (Asamoah et al., 2021 Forson et al., 2020 Pu et
al., 2020). In general, inclusive flotation models based on
relevant flotation variables can be used to better understand
the complexity of the process variables and subsequently
control the future output of the process (Quintanilla et al.,
2021). Over the years, several attempts have been made
to predict froth flotation output in terms of recovery and
grade using approaches like empirical mathematical func-
tions and other unconstrained or unsupervised statistical
approaches (Hodouin, 2011). Recent studies by Gomez-
Flores et al. (2022) and Pu et al. (2020) have recommended
machine learning modelling as more suitable for the empir-
ical modelling of a multivariate unit operation.
Machine learning algorithms can capture complex
nonlinear relationship among high number of flotation
variables for prediction, yet, have received little attention
when it comes to their application in real life industrial data
as most of its success stories are from experimental works
or industrial equipment under controlled conditions (Pu et
al., 2020). Machine learning algorithms including multi-
layer perceptron artificial neural network (ANN), Gaussian
process regression (GPR), support vector machine (SVM),
principal component regression, decision trees and ran-
dom forest (RF) have been applied in predicting metal-
lurgical performance of froth flotation (Ali et al., 2018
Amankwaa-Kyeremeh et al., 2021b Amankwaa-Kyeremeh
et al., 2021d). For instance, Shahbazi et al. (2017) applied
RF and its variable importance measurement in investigat-
ing the effects of particle characteristics and hydrodynamic
conditions, energy dissipation and bubble surface area flux
on flotation rate constant and recovery. Following this, they
predicted flotation rate constant and recovery based on
selected relevant flotation variables. The accuracy of their
predictive models, assessed in terms of coefficient of deter-
mination R2, was 0.96 and 0.97 for flotation rate constant
and recovery, respectively. Work done by Ali et al. (2018)
featured ANN, RF, adaptive neuro-fuzzy inference system
(ANFIS), Mamdani fuzzy logic (MFL), and hybrid neural
fuzzy inference system (HyFIS) in predicting froth ash con-
tent and combustible recovery of fine high ash coal .The
Authors indicated that there was a marginal difference in
terms of performance using R2 performance indicator with
the highest value being 0.92 for MFL. ANN and multi-
variate statistical models were also applied by Nakhaei and
Irannajad (2013) in the estimation of copper grade and
recovery values in flotation column concentrate of a pilot
plant. Their results indicated that ANN gave the most accu-
rate metallurgical performance prediction, outperforming
all the statistical models investigated in their work.
In terms of the application of machine learning algo-
rithms in the prediction of industrial copper flotation
recovery, our previous work (Amankwaa-Kyeremeh et
al., 2021d) happens to be among the very few works that
have been done in the field of mineral processing. GPR
algorithm was used to predict rougher copper recovery
from selected rougher flotation variables which yielded
correlation coefficient (r) values of 0.99, 0.95 and 0.96,
respectively during training, validation and testing phases
(Amankwaa-Kyeremeh et al., 2023). The influence of the
input variables on the recovery was also assessed in our
work. Given the complex nature of industrial processes
and measured data, investigation of multiple algorithms in
finding the optimum data-defined model is recommended.
Can other algorithms, including simple ones, perform bet-
ter than GPR for the considered industrial data? In this
paper, support vector machines, multi-layer perceptron
artificial neural network, linear regression and random
forest algorithms have been investigated against Gaussian
process regression algorithm in predicting copper recovery
in the rougher flotation circuit. The summary of motiva-
tion for these selected algorithms is shown in Table 1. This
research also provides useful information and methodology
for integrating machine learning and data analytical tech-
niques for addressing critical challenges in the mineral pro-
cessing sector and maximising process performance. The
findings can be used in other systems having similar input
and output targets.
RESEARCH METHODOLOGY
The methodologies used for this work are presented in this
section. Specifically, this section captures data acquisition
and pre-processing, model development, theoretical over-
view of predictive algorithms and models performance
assessment. All the algorithms used in this work were car-
ried out using MATLAB R2020a (64-bit version) software.
Data Acquisition and pre-processing
Rougher flotation data for the study was provided by BHP
Olympic Dam, Australia (Amankwaa-Kyeremeh et al.,
2021d). The process variables included the input variables
[throughput, feed grade, froth depth, reagents (xanthate
and frother) dosages, particle size, and air flowrates] and
output variable [copper recovery in the rougher flotation
circuit]. Over 1 million observations of each flotation
dependent on process variables such as feed grade, mineral-
ogy, pulp density, reagent type addition rate, froth depth and
air flow rate, displaying complex interactions (Amankwaa-
Kyeremeh et al., 2021a Dankwah et al., 2022a, b Forson
et al., 2022 Mathe et al., 2000). The process complexity
has warranted froth flotation modelling to predict recovery
and grade (Asamoah et al., 2021 Forson et al., 2020 Pu et
al., 2020). In general, inclusive flotation models based on
relevant flotation variables can be used to better understand
the complexity of the process variables and subsequently
control the future output of the process (Quintanilla et al.,
2021). Over the years, several attempts have been made
to predict froth flotation output in terms of recovery and
grade using approaches like empirical mathematical func-
tions and other unconstrained or unsupervised statistical
approaches (Hodouin, 2011). Recent studies by Gomez-
Flores et al. (2022) and Pu et al. (2020) have recommended
machine learning modelling as more suitable for the empir-
ical modelling of a multivariate unit operation.
Machine learning algorithms can capture complex
nonlinear relationship among high number of flotation
variables for prediction, yet, have received little attention
when it comes to their application in real life industrial data
as most of its success stories are from experimental works
or industrial equipment under controlled conditions (Pu et
al., 2020). Machine learning algorithms including multi-
layer perceptron artificial neural network (ANN), Gaussian
process regression (GPR), support vector machine (SVM),
principal component regression, decision trees and ran-
dom forest (RF) have been applied in predicting metal-
lurgical performance of froth flotation (Ali et al., 2018
Amankwaa-Kyeremeh et al., 2021b Amankwaa-Kyeremeh
et al., 2021d). For instance, Shahbazi et al. (2017) applied
RF and its variable importance measurement in investigat-
ing the effects of particle characteristics and hydrodynamic
conditions, energy dissipation and bubble surface area flux
on flotation rate constant and recovery. Following this, they
predicted flotation rate constant and recovery based on
selected relevant flotation variables. The accuracy of their
predictive models, assessed in terms of coefficient of deter-
mination R2, was 0.96 and 0.97 for flotation rate constant
and recovery, respectively. Work done by Ali et al. (2018)
featured ANN, RF, adaptive neuro-fuzzy inference system
(ANFIS), Mamdani fuzzy logic (MFL), and hybrid neural
fuzzy inference system (HyFIS) in predicting froth ash con-
tent and combustible recovery of fine high ash coal .The
Authors indicated that there was a marginal difference in
terms of performance using R2 performance indicator with
the highest value being 0.92 for MFL. ANN and multi-
variate statistical models were also applied by Nakhaei and
Irannajad (2013) in the estimation of copper grade and
recovery values in flotation column concentrate of a pilot
plant. Their results indicated that ANN gave the most accu-
rate metallurgical performance prediction, outperforming
all the statistical models investigated in their work.
In terms of the application of machine learning algo-
rithms in the prediction of industrial copper flotation
recovery, our previous work (Amankwaa-Kyeremeh et
al., 2021d) happens to be among the very few works that
have been done in the field of mineral processing. GPR
algorithm was used to predict rougher copper recovery
from selected rougher flotation variables which yielded
correlation coefficient (r) values of 0.99, 0.95 and 0.96,
respectively during training, validation and testing phases
(Amankwaa-Kyeremeh et al., 2023). The influence of the
input variables on the recovery was also assessed in our
work. Given the complex nature of industrial processes
and measured data, investigation of multiple algorithms in
finding the optimum data-defined model is recommended.
Can other algorithms, including simple ones, perform bet-
ter than GPR for the considered industrial data? In this
paper, support vector machines, multi-layer perceptron
artificial neural network, linear regression and random
forest algorithms have been investigated against Gaussian
process regression algorithm in predicting copper recovery
in the rougher flotation circuit. The summary of motiva-
tion for these selected algorithms is shown in Table 1. This
research also provides useful information and methodology
for integrating machine learning and data analytical tech-
niques for addressing critical challenges in the mineral pro-
cessing sector and maximising process performance. The
findings can be used in other systems having similar input
and output targets.
RESEARCH METHODOLOGY
The methodologies used for this work are presented in this
section. Specifically, this section captures data acquisition
and pre-processing, model development, theoretical over-
view of predictive algorithms and models performance
assessment. All the algorithms used in this work were car-
ried out using MATLAB R2020a (64-bit version) software.
Data Acquisition and pre-processing
Rougher flotation data for the study was provided by BHP
Olympic Dam, Australia (Amankwaa-Kyeremeh et al.,
2021d). The process variables included the input variables
[throughput, feed grade, froth depth, reagents (xanthate
and frother) dosages, particle size, and air flowrates] and
output variable [copper recovery in the rougher flotation
circuit]. Over 1 million observations of each flotation