XXXI International Mineral Processing Congress 2024 Proceedings/Washington, DC/Sep 29–Oct 3 1037
Primary Crushing Model
The primary crushing model was developed and tuned
using the crusher design testwork data due to the lack of
PSD data for the crushing circuit. The model was developed
using Andersen and Whiten Crushing model (Whiten,
1972 and Andersen and Napier-Munn, 1988).
Primary Grinding—SAG Milling
The variable rates model (Morrell et al., 2001) was uti-
lised to describe the operation of the semi-autogenous mill
(SAG) mill in the flow sheet. The Variable Rates SAG mill
model incorporates the dimensions and operating condi-
tions of the mill, as well as ore-specific hardness parameters
(A, b, and ta) and the throughput of a base case condition.
These parameters are used to determine the machine-spe-
cific breakage rates at various particle sizes. By combining
these breakage rates with a population mass balance model,
the SAG mill performance can be predicted for a given feed
size distribution and ore hardness. The predicted SAG mill
performance includes throughput, volumetric total load,
particle size distribution (PSD), and power draw.
The variables that were used to fine-tune the SAG mill
model were:
• SAG feed %solids
• Drop Weight Index (DWi)—converted to A, b, and
ta values
• SAG feed sizes
• plant throughput
• SAG %critical speed (SAG mill Speed)
Secondary Grinding—Ball Milling
The ball milling operation in the flow sheet was described
using the perfect mixing ball mill model developed by
Whiten (1976). The ‘Perfect Mixing Model’ enables the
scaling of breakage rates from baseline conditions to a vari-
ety of different operating conditions for which performance
is projected.
Classification—Hydrocyclone
The Nageswararao model (1978) was employed to charac-
terise the operation of the hydrocyclone in the flow sheet.
This model combines the cyclone dimensions with the
operating conditions through a series of regression equa-
tions to determine the classification efficiency curve. The
classification efficiency curve is used to predict the size dis-
tributions and flow rates of the two product streams from
the hydrocyclone.
Flotation Circuit—Machine Learning Model
There were no suitable flotation survey or laboratory kinetic
data available to develop phenomenological models for the
flotation circuit. The only flotation data available for mod-
elling were feed and final concentrate assays which were
metallurgically balanced daily. Due to the time constraints
of the project, conducting surveys were not feasible and a
machine learning model was developed using this data to
predict the performance of the overall flotation circuit.
The machine learning model was built using IES
ModelNet package. A total of 1305 daily data were avail-
able from the site spanning 1/2019 to 9/2022 which were
split 80%/10%/10% for training, validation, and testing
respectively. Due to the limited data available, only the flo-
tation throughput, feed P80 size, and feed Cu grade were
used to predict the Cu Recovery. This way, the effects of ore
characteristics and operating conditions in the blasting and
comminution process can be captured by the response of
the flotation model.
Figure 2 depicts parity chart for the Cu recovery model
for the training, validation and test dataset. The figure
Figure 2. The flotation circuit Cu recovery (monthly average)
Primary Crushing Model
The primary crushing model was developed and tuned
using the crusher design testwork data due to the lack of
PSD data for the crushing circuit. The model was developed
using Andersen and Whiten Crushing model (Whiten,
1972 and Andersen and Napier-Munn, 1988).
Primary Grinding—SAG Milling
The variable rates model (Morrell et al., 2001) was uti-
lised to describe the operation of the semi-autogenous mill
(SAG) mill in the flow sheet. The Variable Rates SAG mill
model incorporates the dimensions and operating condi-
tions of the mill, as well as ore-specific hardness parameters
(A, b, and ta) and the throughput of a base case condition.
These parameters are used to determine the machine-spe-
cific breakage rates at various particle sizes. By combining
these breakage rates with a population mass balance model,
the SAG mill performance can be predicted for a given feed
size distribution and ore hardness. The predicted SAG mill
performance includes throughput, volumetric total load,
particle size distribution (PSD), and power draw.
The variables that were used to fine-tune the SAG mill
model were:
• SAG feed %solids
• Drop Weight Index (DWi)—converted to A, b, and
ta values
• SAG feed sizes
• plant throughput
• SAG %critical speed (SAG mill Speed)
Secondary Grinding—Ball Milling
The ball milling operation in the flow sheet was described
using the perfect mixing ball mill model developed by
Whiten (1976). The ‘Perfect Mixing Model’ enables the
scaling of breakage rates from baseline conditions to a vari-
ety of different operating conditions for which performance
is projected.
Classification—Hydrocyclone
The Nageswararao model (1978) was employed to charac-
terise the operation of the hydrocyclone in the flow sheet.
This model combines the cyclone dimensions with the
operating conditions through a series of regression equa-
tions to determine the classification efficiency curve. The
classification efficiency curve is used to predict the size dis-
tributions and flow rates of the two product streams from
the hydrocyclone.
Flotation Circuit—Machine Learning Model
There were no suitable flotation survey or laboratory kinetic
data available to develop phenomenological models for the
flotation circuit. The only flotation data available for mod-
elling were feed and final concentrate assays which were
metallurgically balanced daily. Due to the time constraints
of the project, conducting surveys were not feasible and a
machine learning model was developed using this data to
predict the performance of the overall flotation circuit.
The machine learning model was built using IES
ModelNet package. A total of 1305 daily data were avail-
able from the site spanning 1/2019 to 9/2022 which were
split 80%/10%/10% for training, validation, and testing
respectively. Due to the limited data available, only the flo-
tation throughput, feed P80 size, and feed Cu grade were
used to predict the Cu Recovery. This way, the effects of ore
characteristics and operating conditions in the blasting and
comminution process can be captured by the response of
the flotation model.
Figure 2 depicts parity chart for the Cu recovery model
for the training, validation and test dataset. The figure
Figure 2. The flotation circuit Cu recovery (monthly average)