XXXI International Mineral Processing Congress 2024 Proceedings/Washington, DC/Sep 29–Oct 3 929
Reagent Modelling
To tackle the complexity of capturing the action of reagents
on flotation performance we have taken a Machine Learning
(ML) modelling approach. An explainable grey-box model
is developed to relate feed material properties, cell hydrody-
namic conditions, reagent dosage and circuit performance
(concentrate/tail grade and recoveries). An XGBoost model
(Chen and Guestrin 2016) is built by mixing the client’s
laboratory and process data with hydrodynamic Virtual
Sensor outputs generated by our models described previ-
ously. This combination of first-principles, semi-empirical
and machine-learning techniques is something we call
Scientific AI, where process knowledge conveyed by phys-
ics-based models better inform machine-learning models
allowing for more robust and explainable AI solutions. The
Optimizer finds the combination of reagents (red dot) that
best maximises the Value Driver of that particular circuit
(e.g maximise metal recovery, reduce gangue in the concen-
trate and/or avoid unnecessarily high reagent usage). Value
Drivers are any targets or strategies that directly influence
the profitability, productivity, or sustainability of a mining
operation (Figure 1).
Explainability is also achieved using techniques such as
SHapley Additive exPlanation (SHAP) values, allowing to
visualise and analyse the effect that each model input has on
the model’s output. SHAP values are based on collaborative
game theory and aims to understand the individual contri-
bution of each player (model input) to the final outcome of
the game (model output) (Hart 1989). It has been widely
used to make machine learning models more transparent
and understandable.
Flotation Optimization
Model-based process optimization not only requires real-
time simulation and an optimization engine, but more
importantly, requires a user interface that allows for seam-
less integration with the current workflow of control room
operators. In this section, we describe the components of
the Flotation Optimizer, from the optimization engine to
how the system is integrated with the IT/OT network and
a brief description of the user interface in use.
Optimizer
As previously described, a Digital Process Model is config-
ured and calibrated to mimic the metallurgical behaviour of
a flotation circuit. These circuits can have multiple cells (in
some cases more than 100, including different type cells)
and if the aim is to set airflows and pulp levels for each cell
and reagent rates at different points within the circuit, the
optimization problem becomes quite complex very quickly.
Also, translating process objectives and constraints into an
objective function can be a daunting process. To tackle such
challenges we have taken a reward-based approach, where
simple partial rewards are set for each process variable and
equipment. Examples of these partial rewards are but are
not limited to:
Figure 1. The reagent dosage model, trained with process data and laboratory analytical results, produces surface responses for
the metallurgical response of the circuit in real time. The Optimizer then chooses the best combination of reagents according
to the set Value Driver
Previous Page Next Page

Extracted Text (may have errors)

XXXI International Mineral Processing Congress 2024 Proceedings/Washington, DC/Sep 29–Oct 3 929
Reagent Modelling
To tackle the complexity of capturing the action of reagents
on flotation performance we have taken a Machine Learning
(ML) modelling approach. An explainable grey-box model
is developed to relate feed material properties, cell hydrody-
namic conditions, reagent dosage and circuit performance
(concentrate/tail grade and recoveries). An XGBoost model
(Chen and Guestrin 2016) is built by mixing the client’s
laboratory and process data with hydrodynamic Virtual
Sensor outputs generated by our models described previ-
ously. This combination of first-principles, semi-empirical
and machine-learning techniques is something we call
Scientific AI, where process knowledge conveyed by phys-
ics-based models better inform machine-learning models
allowing for more robust and explainable AI solutions. The
Optimizer finds the combination of reagents (red dot) that
best maximises the Value Driver of that particular circuit
(e.g maximise metal recovery, reduce gangue in the concen-
trate and/or avoid unnecessarily high reagent usage). Value
Drivers are any targets or strategies that directly influence
the profitability, productivity, or sustainability of a mining
operation (Figure 1).
Explainability is also achieved using techniques such as
SHapley Additive exPlanation (SHAP) values, allowing to
visualise and analyse the effect that each model input has on
the model’s output. SHAP values are based on collaborative
game theory and aims to understand the individual contri-
bution of each player (model input) to the final outcome of
the game (model output) (Hart 1989). It has been widely
used to make machine learning models more transparent
and understandable.
Flotation Optimization
Model-based process optimization not only requires real-
time simulation and an optimization engine, but more
importantly, requires a user interface that allows for seam-
less integration with the current workflow of control room
operators. In this section, we describe the components of
the Flotation Optimizer, from the optimization engine to
how the system is integrated with the IT/OT network and
a brief description of the user interface in use.
Optimizer
As previously described, a Digital Process Model is config-
ured and calibrated to mimic the metallurgical behaviour of
a flotation circuit. These circuits can have multiple cells (in
some cases more than 100, including different type cells)
and if the aim is to set airflows and pulp levels for each cell
and reagent rates at different points within the circuit, the
optimization problem becomes quite complex very quickly.
Also, translating process objectives and constraints into an
objective function can be a daunting process. To tackle such
challenges we have taken a reward-based approach, where
simple partial rewards are set for each process variable and
equipment. Examples of these partial rewards are but are
not limited to:
Figure 1. The reagent dosage model, trained with process data and laboratory analytical results, produces surface responses for
the metallurgical response of the circuit in real time. The Optimizer then chooses the best combination of reagents according
to the set Value Driver

Help

loading